The Valossa Video Recognition API (which is the endpoint for the Video Analysis function and was formerly known as the Core API) is a REST API for automatic video content analysis and metadata creation.

This Video Analysis API provides access to a broad set of high-quality audiovisual AI Features for recognition and analysis, and returns analysis results with details and associated video time segments in the Valossa Metadata format.

Valossa AI is constructed of a set of AI Features (formerly know as capabilities), which can be unlocked in various Subscription Options. Subscriptions enable a set of AI feature configurations for a specific function, such as providing Standard Metadata for describing video details in full, or executing Face Expression and Voice Sentiment analysis. (Read further for more details on Subscriptions.)

Each AI Feature is specialized in detecting and recognizing a set of concepts that define the semantic output to describe a content entity. Detections are grouped and concatenated in JSON-format as Valossa Metadata, the main downloadable output format being called Valossa Core metadata. There are also additional output formats for special, separately downloadable results such as bounding boxes for faces or objects, and also SRT files containing the speech-to-text results from the audio track of the videos.

Video AI features (available in different Video Analysis Subscription configurations)

Faces (gender, face groups, face bounding boxes and detected similarity to persons in your custom face gallery created with the Valossa Training API)
Facial expressions V1 (Theatrical emotions) and Facial expressions V2 (Theatrical and Mild emotions) (for the list of emotional states, see: emotions from faces)
Visual tags (visual entities like objects, animals, scenes and styles, grouped into powerful Tag Categories)
Audio tags (music styles and musical moods, instruments, sounds of human actions and animals and machines etc., sound effects)
Speech-to-text (speech-to-text results in supported languages, including NLP i.e. natural language processing to provide topical keywords and named entities from speech, with appearance time codes)
Speech sentiment (positive, neutral or negative sentiment from speech)
Visual tags for Inappropriate Content (visual nudity, sexual scenes, violence, injuries, weapons, substance use, accidents and disasters; see Tag Categories to review all possibilities)
Audio offensive words from audio
Video OCR (optical character recognition, that is, recognition of text segments in video frames)
Dominant scene colors
Video shot segmentation
Video topics from audiovisual content, including IAB categorization

Valossa Subscription Options (contact sales to order and activate):

Standard Metadata: Extract a comprehensive, contextual, scene-level content metadata of your videos. Capabilities included:
- Face detection and recognition
- Face training to create user specific face identities
- Visual tags
- Audio tags
- Speech-to-text
- Visual tags for inappropriate content
- Video OCR
- Dominant scene colors
- Video Topics and Tag Categories
- (Optional Add-on) Facial expressions and Voice sentiment
Face Analysis with Expressions and Voice Sentiment: Focus on face emotions and video sentiment analysis. Capabilities included:
- Face detection and recognition
- Face training to create user specific face identities
- Facial expressions
- Speech-to-text
- Speech sentiment
- Voice emotion
Visual Video Moderation: Focus on detecting inappropriate content from visuals and audio. Capabilities included:
- Visual tags (for compliance objects such as guns, needles)
- Visual tags for inappropriate content (nudity, violence etc.)
- (Add-on, enabled by default) Audio offensive words from audio, and sounds such as gunshots
Autopreview for Automated Video Highlights Production: Extract highlights and create promotions automatically from your videos. Capabilities included:
- Valossa Autopreview for automated highlights creation

Additional capabilities available in customized setups (contact sales to request more information). For example:

Logo recognition (global consumer brands)
Football (soccer) events (alpha version)

Detections are also provided in different practical groupings in the metadata: grouped by detection type (best concepts first, and including time-coded occurrences of the concept when applicable) and grouped by second ("What is detected at 00:45 in the video?"). See explanation of the practical detection groupings

Valossa Video Recognition API overview

The REST API is invoked using HTTP (HTTPS) requests. You can also assign new video analysis jobs to the API on the easy API call page. The responses from the REST API are in the JSON format. The Valossa Report tool helps you to visualize the results. Please use the secure HTTPS transport in all API calls to protect your valuable data and confidential API key: unencrypted HTTP transport is not allowed by the API.

Get your API key from under "My account" - "API keys" in Valossa Portal. If you have several applications, you may request a different API key for each of them; for this, contact Valossa service personnel.

Note: As the administrator user of your organization, you can create new users under "My account" - "Manage users". If your organization has several people who need to use the Portal, you should add them manually in "Manage users", so they are all mapped to your organization and may view analysis results (if you give the rights to the users in Portal) and post new jobs (if you give the rights). The permissions are mappings between users and API keys ("this user has read-write access to this API key so she can both view results and make new job requests"), so please configure the permissions understanding this; the API key permissions per user can be edited in "Manage users". For your company/organization, you must have only one customer account (created by the Valossa sales staff), but there can be multiple users under the customer account!

The API consists of 6 different functions:

new_job [HTTP POST]
This function is used to create a new video analysis job in our system. The job (including e.g. the URL of the video file to be analyzed) is defined by using a JSON formatted data structure that is included as the body of the HTTP POST request. This function returns the job_id (UUID value) of the created job. The job_id is used after this as the identifier when querying the status and the results of the job. The video recognition capabilities to use in this job are bound to the API key that you specified for the job.
job_status [HTTP GET]
This function is used to query (poll) the status of a specific job, based on its job_id.
job_results [HTTP GET]
This function is used to fetch the resulting metadata JSON of a finished analysis job identified by its job_id.
list_jobs [HTTP GET]
This function lists all the jobs for a given API key.
cancel_job [HTTP POST]
This function cancels a job, if it is in a cancellable state.
delete_job [HTTP POST]
This function deletes a job and the assets (video file etc.) associated with it.

You can conveniently monitor the status of your jobs in Valossa Portal. There you can also call the new_job function of the API with an easy API request generator.

Your API key is shown in Valossa Portal on the subscriptions page. Keep the key confidential.

Please note regarding speech analysis:

If you already have the speech transcript of your video in the SRT format (for example the subtitles of your movie), please specify the transcript URL in the request, along with the video URL. The transcript content will be analyzed, and the detected concepts will be included in the "transcript" part of the metadata JSON.
Your existing transcript is, obviously, a more reliable source for speech information than audio analysis. So, if you have the transcript, please use it – it’s a valuable asset!
In late 2019, a new feature was introduced. Now you can run an analysis job where the input is only a transcript (SRT) without a video file. For this, you need to make some special configurations. See instructions for running a transcript-only analysis
Audio keyword detection and audio speech-to-text will be performed only if you did not provide the SRT transcript (however, providing or omitting the SRT transcript does not affect the "audio.context" detections). If you provided your own transcript, the results derived from it are available in the "transcript.*" detections.
The audio-related metadata generated by us contains, among other things, the actual speech-to-text results. For accessing these speech-to-text results, please see the instructions related to the detection types "audio.speech" and "audio.speech_detailed" as well as the separately downloadable SRT-formatted or VTT-formatted speech-to-text results. In addition, we provide you a uniquely descriptive set of keywords extracted from the speech. Whether the source of speech information is audio itself or your transcript file, the output format of the detected keywords is similar in the metadata (of course, there is a difference at least in the detection type identifiers: the audio.keyword.* detection types vs. the transcript.keyword.* detection types).

Input data formats

Supported video formats: we support most typical video formats, including but not limited to MP4, MPEG, AVI, FLV, WebM, with various codecs. Currently, we cannot provide a fixed list of supported formats and codecs, but for example MP4 with the H.264 codec works.

Video file size limit: 5GB per input video file.

Video duration limit: 3 hours of playback time per input video file.

Video vertical resolution limit: 1080 pixels.

Currently, the supported languages for speech-based detections are English, French, German, Spanish, Finnish, Italian, Brazilian Portuguese, European Portuguese, Swedish and Dutch. By default, speech is analyzed as English language. See more information about language selection.

If the video file contains several video streams, only the first one is analyzed.

If the video file contains several audio streams, only the first one is analyzed. (Please note that audio keyword detection and audio speech-to-text will be performed only if you did not provide your own SRT-based speech transcript; however, providing or omitting the SRT transcript does not affect the audio.context detections.) The audio stream can be either mono or stereo.

Supported transcript format: SRT.

File size limit: 5MB per input SRT file.

Currently, the only supported transcript language is English.

Usage examples

Creating a new video analysis job

You must pay for the video analysis. How to pay? How to gain access to Valossa AI? Start a subscription by contacting sales and pay the invoices according to the agreement.

Which of the many video recognition capabilities of Valossa will be used for the specific job, depends on the configuration of the API key used in the new_job request. The subscription you agreed with Valossa determines the video recognition capabilities of your API key(s). An example of subscription types is the Standard Metadata subscription, containing a wide range of audiovisual detection types. View your API keys on the subscriptions page in Valossa Portal. Keep the API keys confidential, otherwise someone can impersonate you or your application with a leaked key.

How to create a new video analysis job using the REST API? Send an HTTP POST to the URL:

https://api.valossa.com/core/1.0/new_job

Example new_job request body in JSON format:

{
  "api_key" : "kkkkkkkkkkkkkkkkkkkk",
  "media": {
    "title": "The Action Movie",
    "description": "Armed with a knife, Jack Jackson faces countless dangers in the jungle.",
    "video": {
      "url": "https://example.com/content/Action_Movie_1989.mp4"
    },
    "transcript": {
      "url": "https://example.com/content/actionmovie.srt"
    },
    "customer_media_info": {
      "id": "469011911002"
    },
    "language": "en-US"
  }
}

There are two different ingestion methods for your video file: download and upload. Downloading involves a URL that points to your downloadable video file, and you use the download URL directly in your new_job request. Uploading involves first the use of the upload functionality of the Valossa Video Recognition API. After the upload has been completed, you use the resulting valossaupload:// URL in your new_job request to refer to the specific uploaded file. See instructions for video file uploading

In the download video ingestion method, the video URL and transcript URL can be either http:// or https:// or s3:// based. If the URL is s3:// based, you should first communicate with us to ensure that our system has read access to your S3 bucket in AWS (Amazon Web Services).

Whether you used the download or upload video ingestion method, the video URL is mandatory in the new_job request. The URL must directly point to a downloadable video file (in which case, our system will download the file from your system) or it must be a valossaupload:// URL of your uploaded file.

The transcript URL is optional – but recommended, because an existing SRT transcript is a more reliable source of speech information than audio analysis. The URL must directly point to a downloadable SRT transcript file. Our system will download the file from your system.

The title is optional – but recommended: a human-readable title makes it easy for you to identify the video on the results page of Valossa Portal, and will also be included in the metadata file.

The description is optional. Description is any freetext, in English, that describes the video.

If title and/or description are provided in the call, the text in them will also be analyzed, and the detected concepts will be included in the analysis results (the "external" concepts in the metadata JSON).

The customer media info is optional. If you provide a customer media ID in the "id" field inside the "customer_media_info" field, you may use the customer media ID (a string from your own content management system) to refer to the specific job in the subsequent API calls, replacing the "job_id" parameter with a "customer_media_id" parameter in your calls. Note: Our system will NOT ensure that the customer media ID is unique across all jobs. Duplicate IDs will be accepted in new_job calls. It is the responsibility of your system to use unique customer media IDs, if your application logic requires customer media IDs to be unique. If you use duplicate customer media IDs, then the latest inserted job with the specific customer media ID will be picked when you use the "customer_media_id" parameter in the subsequent API calls.

The language is optional. It specifies the language model to be used for analyzing the speech in the audio track of your video. The allowed values are "de-DE" (German), "en-US" (US English), "es-ES" (Spanish), "fi-FI" (Finnish), "fr-FR" (French), "it-IT" (Italian), "pt-BR" (Brazilian Portuguese), "pt-PT" (European Portuguese), "sv-SE" (Swedish) and "nl-NL" (Dutch). More languages will be supported in the future. If the language parameter is not given, the default "en-US" will be used so the speech in the video is assumed to be in US English.

Please note that for other languages than US English, the following exceptions apply.

Audio-based detections of named entities ("audio.keyword.name.person", "audio.keyword.name.location", "audio.keyword.name.organization", "audio.keyword.name.general" detections) and offensive words ("explicit_content.audio.offensive" detections) are available, but audio-based detections of novelty words ("audio.keyword.novelty_word" detections) are not available.
- However, for the French language, "audio.keyword.novelty_word" detections are available (as English translations).
Audio-based IAB categories are not available.

The language-specific details are subject to change in the future.

If the analysis is technically successful (i.e. if the job reaches the "finished" state), the job will be recorded as part of the used volume on your ongoing subscription period. No subscription yet? Contact sales.

Here is an example new_job request body with only the mandatory fields present:

{
  "api_key": "kkkkkkkkkkkkkkkkkkkk",
  "media": {
    "video": {
      "url": "https://example.com/my-car-vid.mpg"
    }
  }
}

Here is an example new_job request body with the specification to use a non-default, self-created face gallery, so the faces in that gallery will be used for identifying the persons in the video:

{
  "api_key": "kkkkkkkkkkkkkkkkkkkk",
  "media": {
    "video": {
      "url": "https://example.com/my-car-vid.mpg"
    }
  },
  "analysis_parameters": {
    "face_galleries": {
      "custom_gallery": {
        "id": "468a2b70-3b55-46f8-b209-8ad2fcabd5c8"
      }
    }
  }
}

If you have a default face gallery and want to use it in your video analysis job, just leave out the gallery selection in the new_job call. The default face gallery will be implicitly selected, when no other gallery is explicitly selected.

The response of a successful new_job call always includes the job_id of the created job.

Example response in an HTTP 200 OK message:

{
  "job_id": "6faefb7f-e468-43f6-988c-ddcfb315d958"
}

Jobs are identified by UUIDs, which appear in "job_id" fields in various messages. Your script that calls the API must, of course, save the job_id from the new_job response in order to be able to query for the status and results later.

Example test call with Curl on the command line, if your test request JSON is in a file created by you:

curl --header "Content-Type:application/json" -X POST -d @your_request.json https://api.valossa.com/core/1.0/new_job

If you want a HTTP POST callback and/or email notification when your video analysis job reaches an end state, you may specify one or both of those in the new_job request. The HTTP POST callback mechanism in our system expects your system to send a 200 OK response for the request (callback) initiated by our system. The request will be re-tried one time by our system, if the first attempt to access your specified callback URL returns a non-200 code from your system or times out. Due to the possibility of network problems and other reasons, you should not rely on the HTTP POST callback to be received by your system. In any case, whether the HTTP POST callback event was received or not, your system can always check the status of the job using the job_status function in the REST API. The email notification will be sent to those users that have the permission to view job results for the chosen API key.

Example of a job request with a HTTP POST callback:

{
  "api_key": "kkkkkkkkkkkkkkkkkkkk",
  "callback": {
    "url": "https://example.com/your_callback_endpoint"
  },
  "media": {
    "title": "Lizards dancing",
    "video": {
      "url": "https://example.com/lizards_dancing.mkv"
    }
  }
}

The HTTP POST callback message is formatted as JSON, and contains the job ID in the "job_id" field and the reached end status of the job in the "status" field. It also contains the customer media ID in the "customer_media_id" field, if you had given a customer media ID for the job. Here is an example of a HTTP POST callback message body:

{
  "job_id": "ad48de9c-982e-411d-93a5-d665d30c2e92",
  "status": "finished"
}

Example of a job request with an email notification specified:

{
  "api_key": "kkkkkkkkkkkkkkkkkkkk",
  "email_notification": {
    "to_group": "users_with_access_to_api_key_results"
  },
  "media": {
    "title": "Lizards dancing",
    "video": {
      "url": "https://example.com/lizards_dancing.mkv"
    }
  }
}

The generated email notification message is intended for a human recipient. So, unlike the HTTP POST callback, the email notification message is not intended for machine parsing.

How to upload a video file using the API?

There are 3 different API requests that make it easy to upload your file. Because videos often are huge, the files must be uploaded in chunks.

First, you initialize the upload. Send an HTTP POST to the URL:

https://api.valossa.com/core/1.0/initialize_file_upload

You must specify the size of the file to be uploaded (in bytes), and the response will tell you the chunk size (in bytes) that you must use when sending your file chunks. Save the upload ID (from the "upload_id" field of the response) in order to be able to refer to the same upload action in the subsequent requests.

Curl example:

curl -F "api_key=kkkkkkkkkkkkkkkkkkkk" -F "file_size_bytes=12740839" https://api.valossa.com/core/1.0/initialize_file_upload

Example response in an HTTP 200 OK message:

{
  "upload_id": "7c62bc7b-e143-4a81-aa83-a7eb0ec37077",
  "file_chunk_size_bytes": 4194304
}

The Content-Type header of the initialize_file_upload request must be "multipart/form-data". If you use Curl and its "-F" option, Curl will set this Content-Type as default and will also use POST as the HTTP method in the request.

Next, upload each chunk of your video file. The size of each chunk must be exactly the number of bytes indicated in the "file_chunk_size_bytes" field in the response you received for your initialize_file_upload request. However, the last chunk of the file may have a different size; this is natural, because the total size of the file usually is not an exact multiple of the chunk size.

How to split your video file into chunks? In Linux-style operating systems, the "split" command is suitable for the task. We recommend using the "-d" option of the "split" command, because then the chunks will be named so that the chunk index in the chunk file names is a number sequence rather than a letter sequence that is the default. The number-based indexing is probably easier to use in your own helper scripts.

Split example:

split -d --bytes=4194304 supervideo.mp4 supervideo.mp4.chunk

The above command creates the files supervideo.mp4.chunk00, supervideo.mp4.chunk01, supervideo.mp4.chunk02 with a size of 4194304 bytes and the file supervideo.mp4.chunk03 with a size of 157927 bytes.

When uploading file chunks, you must specify the upload ID and chunk index. The chunk index is an integer, and the first chunk of the file has the index 0. The temporal order of the send_file_chunk requests with the different chunk indexes does not need to be 0, 1, 2... but you must specify the correct index for each chunk, so the uploaded file will be correct when the chunks are reassembled into the full file in our system. Of course, probably it is easiest for you to use a helper script that just sends the chunks in the order 0, 1, 2..., increasing the value of chunk_index in a loop.

Send an HTTP POST to the URL:

https://api.valossa.com/core/1.0/send_file_chunk

Curl example for uploading the first chunk:

curl -F "api_key=kkkkkkkkkkkkkkkkkkkk" -F "upload_id=7c62bc7b-e143-4a81-aa83-a7eb0ec37077" -F "chunk_index=0" -F "file_data=@supervideo.mp4.chunk00" https://api.valossa.com/core/1.0/send_file_chunk

The Content-Type header of the send_file_chunk request must be "multipart/form-data".

Example response in an HTTP 200 OK message:

{}

When all chunks have been uploaded, you need to finalize the upload, referring to the correct upload ID. Send an HTTP POST to the URL:

https://api.valossa.com/core/1.0/finalize_file_upload

Curl example:

curl -F "api_key=kkkkkkkkkkkkkkkkkkkk" -F "upload_id=7c62bc7b-e143-4a81-aa83-a7eb0ec37077" https://api.valossa.com/core/1.0/finalize_file_upload

The Content-Type header of the finalize_file_upload request must be "multipart/form-data".

Example response in an HTTP 200 OK message:

{
  "uploaded_file_url": "valossaupload://7c62bc7b-e143-4a81-aa83-a7eb0ec37077"
}

Save the valossaupload:// URL from the "uploaded_file_url" field of the response, so you can use it in your new_job request for that video.

For an entirely manual upload of a file in a graphical user environment, use the Analyze page in Valossa Portal.

Error responses from the API

The following pertains to the HTTP error responses, which are returned immediately for your API call if your request was malformed or missing mandatory fields. In other words, the following does not pertain to the separate HTTP callback messages, which were discussed above. (Callback events are not even generated for the errors that are returned immediately in the HTTP error response of an API call.)

Error responses from the API calls (new_job calls or any other calls) contain an error message, and can be automatically separated from 200 OK responses, because error responses are sent along with an HTTP error code (non-200). Error responses are also formatted as JSON, and they contain an "errors" array, where one or more errors are listed with the corresponding error messages.

Example error response in an HTTP 400 message:

{
  "errors": [
    {
      "message": "Invalid API key"
    }
  ]
}

Getting status of a single job

The status of a single analysis job is polled using HTTP GET.

Example request:

https://api.valossa.com/core/1.0/job_status?api_key=kkkkkkkkkkkkkkkkkkkk&job_id=6faefb7f-e468-43f6-988c-ddcfb315d958

Example response in an HTTP 200 OK message:

{
  "status": "processing",
  "media_transfer_status": "finished",
  "details": null,
  "poll_again_after_seconds": 600
}

Possible values for the "status" field: "queued", "on_hold", "preparing_analysis", "processing", "finished", and "error". More status values may be introduced in the future.

If the job status is "error", something went wrong during the analysis process. If there is an explanation of the error in the "details" field, please see if the cause of the error is something you can fix for yourself (such as a non-video file in the video URL of the job request). Otherwise, contact us in order to resolve the issue.

If the job status is "queued" or "processing", you should poll the status again after some time.

If the job status is "finished", you can fetch the job results using the job_results function.

The "details" field may contain some additional details about the status of the job.

The "media_transfer_status" field indicates whether the media to be analyzed has been transferred from your system to our system. Possible values for the "media_transfer_status" field: "queued", "downloading", "finished" and "error". If "media_transfer_status" is "finished", your video (and the transcript if you provided it) have been successfully transferred to our system.

The value in "poll_again_after_seconds" is just a suggestion about when you should poll the job status again (expressed as seconds to wait after the current job_status request).

If there was a problem with the job_status query itself, the error will be indicated in an HTTP non-200 response with a JSON body, similar to the error responses of the new_job function.

Getting the results of a job

After a job has been finished, the resulting video metadata can be fetched using HTTP GET.

Example request:

https://api.valossa.com/core/1.0/job_results?api_key=kkkkkkkkkkkkkkkkkkkk&job_id=6faefb7f-e468-43f6-988c-ddcfb315d958

Response data is in the JSON format. For more details, see chapter "Output metadata JSON format".

Save the metadata and use it from your own storage disk or database for your easy and quick access. We will not necessarily store the results perpetually.

If there was a problem with the job_results query itself, the error will be indicated in an HTTP non-200 response with a JSON body, similar to the error responses of the new_job function.

Want to search for specific recognized things among your accumulated video analysis job results, by typing search words such as "airplane" or "Emilia Clarke"? Please try the easy-to-use Search functionality in Valossa Portal.

Listing all your jobs and statuses of all your jobs

Convenience function for listing all your jobs, optionally with also their job statuses (optional parameter "show_status" with the value "true"), using HTTP GET:

Example request:

https://api.valossa.com/core/1.0/list_jobs?api_key=kkkkkkkkkkkkkkkkkkkk&show_status=true

Example response in an HTTP 200 OK message:

{"jobs": [
  {
    "job_id": "6faefb7f-e468-43f6-988c-ddcfb315d958",
    "job_status":
    {
      "status": "finished",
      "media_transfer_status": "finished",
      "details": null,
      "poll_again_after_seconds": null
    }
  },{
    "job_id": "36119563-4b3f-44c9-83c6-b30bf69c1d2e",
    "customer_media_id": "M4070117",
    "job_status":
    {
      "status": "processing",
      "media_transfer_status": "finished",
      "details": null,
      "poll_again_after_seconds": 600
    }
  }
]}

If you had given a customer media ID when creating the job, the "customer_media_id" field exists and contains the customer media ID value.

Showing video titles and other media information in the job listing is often useful. This can be done by using the optional GET parameter "show_media_info" with the value "true". Example request:

https://api.valossa.com/core/1.0/list_jobs?api_key=kkkkkkkkkkkkkkkkkkkk&show_status=true&show_media_info=true

Example response in an HTTP 200 OK message:

{"jobs": [
  {
    "job_id": "36119563-4b3f-44c9-83c6-b30bf69c1d2e",
    "customer_media_id": "M4070117",
    "job_status":
    {
      "status": "finished",
      "media_transfer_status": "finished",
      "details": null,
      "poll_again_after_seconds": null,
      "media_info":
      {
        "title": "Birds clip #22",
        "description": "Birds having a bath",
        "video":
        {
          "url": "https://example.com/contentrepository/project1/aabhk-gg4rt-gi5aq-jjv6t/birds_22_original.mp4"
        }
      }
    }
  },{
    "job_id": "6faefb7f-e468-43f6-988c-ddcfb315d958",
    "job_status":
    {
      "status": "finished",
      "media_transfer_status": "finished",
      "details": null,
      "poll_again_after_seconds": null,
      "media_info":
      {
        "video":
        {
          "url": "https://example.com/my-car-vid.mpg"
        }
      }
    }
  }
]}

By adding the optional GET parameter "n_jobs" to the request (example: n_jobs=500), you can control how many of your jobs will be listed if your job list is long. The default is 200. The maximum possible value for "n_jobs" is 25000.

If there was a problem with the list_jobs query itself, the error will be indicated in an HTTP non-200 response with a JSON body, similar to the error responses of the new_job function.

Cancel a job

Cancel a job by sending a HTTP POST to the URL:

https://api.valossa.com/core/1.0/cancel_job

Example cancel_job request body:

{
  "api_key": "kkkkkkkkkkkkkkkkkkkk",
  "job_id": "be305b1e-3671-45b1-af88-1f052db3d1bb"
}

Example response in an HTTP 200 OK message:

{
  "job_status": "canceled"
}

The job must be in a cancellable state for this function to succeed. For example, a finished job is not cancellable.

If there was a problem with the cancel_job query itself, the error will be indicated in an HTTP non-200 response with a JSON body, similar to the error responses of the new_job function.

Delete a job

Delete a job by sending a HTTP POST to the URL:

https://api.valossa.com/core/1.0/delete_job

Example delete_job request body:

{
  "api_key": "kkkkkkkkkkkkkkkkkkkk",
  "job_id": "f3cd3108-444a-4c06-84a6-730ac231e431"
}

Example response in an HTTP 200 OK message:

{}

Deleting a job will remove it from the set of your existing jobs. When a job is deleted, the assets (video file etc.) associated with the job are also deleted. There may be some delay between completing the delete_job query and the purging of all assets from the storage systems in the background.

If there was a problem with the delete_job query itself, the error will be indicated in an HTTP non-200 response with a JSON body, similar to the error responses of the new_job function.

Valossa Training API overview

The Valossa Training API is part of those Valossa AI subscriptions that contain faces-related functionality. In addition to REST API access, the functionalities of the Valossa Training API can be accessed using a graphical user interface in Valossa Portal.

Using the Valossa Training API, you can train the system to detect custom faces. The custom faces will be detected in those videos that you analyze after the training. By training your custom faces, you acknowledge and accept the fact that using custom-trained faces may cause some additional delays in the processing of your video analysis jobs.

The detected face identities will appear in the "similar_to" fields of the "human.face" detections in Valossa Core metadata. Your API key(s) that work for creating new analysis jobs with the Valossa Video Recognition API will also work for face training with the Valossa Training API, if faces-related functionality is included in your active Valossa AI subscription.

How to create your custom face gallery? There are two ways:

Your default gallery, and the ID (a UUID) of that gallery, are created implicitly when you start training custom faces if you don't explicitly specify a gallery.
You can explicitly create a non-default gallery. The ID (a UUID) of that gallery is provided by the gallery creation operation. To a self-created non-default gallery, you can give a name (unlike with the default gallery) and you can have several non-default galleries (as opposed to the single default gallery). A typical reason for having several named, non-default galleries is that each of your own customers requires a separate face gallery in your application that interfaces with Valossa AI. When using separate galleries, the faces of your customer A will not be used for the video analysis jobs related to your customer B, if you correctly specify the face gallery to use when creating the video analysis jobs.

From these Curl-based request-and-response examples it is easy to modify REST API calls for use in your application. Just like with the Video Recognition API, the HTTP response code 200 indicates a successful operation, and a non-200 code indicates error (an error message is provided in that case). The response body is in the JSON format.

As you can see from the examples, any "read data" requests use the HTTP GET method, while any "write data" or "erase data" requests use the HTTP POST method.

The trained faces can be used in your subsequent video analysis jobs to create the "similar_to" items in the "human.face" detections with the correct personal identities. Of course, the faces to use are picked from the gallery specified in the new_job request. If no gallery is specified, then your default face gallery is used for the video analysis job. A non-default gallery must be explicitly specified in the new_job request in order to be used in the video analysis job.

Uploading images for training

Adding sample images for a face has been designed to work with both, file uploads and file downloads. Thus, the file reference mechanism used in the add_face_image request of the Valossa Training API uses an easy URL-based syntax for both file input styles. Currently, only uploads are supported, but download support will be added in the future. Download, obviously, means that our system downloads the image file from an HTTP(S) URL provided by your system. Uploaded files each get assigned a valossaupload:// URL that uniquely identifies the successfully received file that resides in our storage system.

When you upload a file, first use the REST function upload_image to move the file content. Then, you will use the valossaupload:// URL of the correct file when referring to the file in the add_face_image request in the actual face identity training.

Before using the service in a way that makes you a "processor" and/or "controller" of the personal data of EU-resident natural persons, you are required to make sure that your actions are compliant with the General Data Protection Regulation. See the Terms and Conditions of Valossa services

An image must be in the JPG or PNG format. The maximum image file size is 8MB. The maximum width of an image is 4096 pixels. The maximum height of an image is 4096 pixels.

At least 10 different sample images of each face, photographed from different angles etc., should be given in order to get good training results. The more images the better. Training may in some cases work even with only a few images, but the results are better with more samples: a lot of clear, diverse, high-quality images of the face to be trained.

Send an HTTP POST to the URL:

https://api.valossa.com/training/1.0/upload_image

Curl example:

curl -F "image_data=@ricky_1.jpg" -F "api_key=kkkkkkkkkkkkkkkkkkkk" https://api.valossa.com/training/1.0/upload_image

The upload_image call is similar regardless of the gallery type (default vs. non-default), because this request does not operate on a gallery; it just uploads an image for further use that probably involves a gallery.

Example response in an HTTP 200 OK message:

{
  "uploaded_file_url": "valossaupload://ff357efe-1086-427d-b90c-1d1887fb1017"
}

The Content-Type header of the file upload request must be "multipart/form-data". If you use Curl and its "-F" option, Curl will set this Content-Type as default and will also use POST as the HTTP method in the request. There must be one file per upload_image request.

Note! As you can see from the Curl request example above, the API key must be sent as a form parameter (not URL parameter). This is quite natural, taking into account that the Content-Type of the request is "multipart/form-data".

Training face identities and managing the trained identities

All the POST-based REST functions listed below accept a JSON-formatted input string, which contains the parameters of the specific function. The GET-based REST functions read their parameters from the request URL.

Create a new face gallery

You may decide to use your default face gallery that is created implicitly when you start training faces without specifying a face gallery. In that case, please skip face gallery creation. Otherwise, please explicitly create your non-default face gallery based on the following instructions.

Send an HTTP POST to the URL:

https://api.valossa.com/training/1.0/create_face_gallery

Curl example:

curl --header "Content-Type:application/json" -X POST -d '{"api_key":"kkkkkkkkkkkkkkkkkkkk", "gallery":{"name":"My Special Faces"}}' https://api.valossa.com/training/1.0/create_face_gallery

Example response in an HTTP 200 OK message:

{
  "gallery_id": "468a2b70-3b55-46f8-b209-8ad2fcabd5c8"
}

Save the gallery ID locally. You will need it when you add face identities to the gallery or when you do any other operations with the specific gallery.

The maximum length of the "name" parameter of a face gallery is 1024 characters.

Update face gallery

Send an HTTP POST to the URL:

https://api.valossa.com/training/1.0/update_face_gallery

The "updates" structure contains the face gallery parameters to update. Currently the only allowed parameter is "name". The data type for this value is string. The maximum length of the value of "name" is 1024 characters.

Curl example:

curl --header "Content-Type:application/json" -X POST -d '{"api_key":"kkkkkkkkkkkkkkkkkkkk", "gallery":{"id":"468a2b70-3b55-46f8-b209-8ad2fcabd5c8", "updates":{"name":"My Extra-Special Faces"}}}' https://api.valossa.com/training/1.0/update_face_gallery

Example response in an HTTP 200 OK message:

{}

Only a non-default face gallery can be updated. The default face gallery cannot be modified by this request. If you need named galleries, you should be using non-default, explicitly created face galleries.

List existing face galleries

Send an HTTP GET to the URL:

https://api.valossa.com/training/1.0/list_face_galleries

Curl example:

curl 'https://api.valossa.com/training/1.0/list_face_galleries?api_key=kkkkkkkkkkkkkkkkkkkk'

Example response in an HTTP 200 OK message:

{
  "face_galleries": [
    {
      "id": "cc003b8e-6c67-491b-95c2-9155dc894549"
    },
    {
      "id": "468a2b70-3b55-46f8-b209-8ad2fcabd5c8"
    }
  ]
}

It is also possible to list existing face galleries with details.

Curl example:

curl 'https://api.valossa.com/training/1.0/list_face_galleries?api_key=kkkkkkkkkkkkkkkkkkkk&show_details=true'

Example response in an HTTP 200 OK message:

{
  "face_galleries": [
    {
      "id": "cc003b8e-6c67-491b-95c2-9155dc894549",
      "created_at": "2018-01-27 10:03:29",
      "is_default": true
    },
    {
      "id": "468a2b70-3b55-46f8-b209-8ad2fcabd5c8",
      "name": "My Extra-Special Faces",
      "created_at": "2018-10-18 12:22:47",
      "is_default": false
    }
  ]
}

Create a new face identity

The string fields "name" and "gender" are optional. We recommend setting at least the name, because a nameless face identity might cause confusion for you later on (however, it is perfectly acceptable to have a nameless face identity, if your application logic requires creating such an identity). The maximum length of the value of "name" is 1024 characters. The gender is "male" or "female". The response contains the unique identifier of the face identity (person).

Send an HTTP POST to the URL:

https://api.valossa.com/training/1.0/create_face_identity

Curl example if you are using your default gallery:

curl --header "Content-Type:application/json" -X POST -d '{"api_key":"kkkkkkkkkkkkkkkkkkkk", "face":{"name":"Ricky Rickson", "gender":"male"}}' https://api.valossa.com/training/1.0/create_face_identity

Curl example if you are using a non-default gallery:

curl --header "Content-Type:application/json" -X POST -d '{"api_key":"kkkkkkkkkkkkkkkkkkkk", "gallery":{"id":"468a2b70-3b55-46f8-b209-8ad2fcabd5c8"}, "face":{"name":"Ricky Rickson", "gender":"male"}}' https://api.valossa.com/training/1.0/create_face_identity

Example response in an HTTP 200 OK message:

{
  "face_id": "bb254a82-08d6-4498-9ddb-3de4c88f1f66"
}

Save the face ID locally. You will need it when you add images for the face or when you do any other operations with the specific face identity.

Add face images to a specific face identity

Referring to your previously uploaded files, you need to add the correct files to a specific existing face identity, one image file per add_face_image request. The response contains a unique identifier of the processed, accepted training image, from which a sample face has been detected. You need the ID later, if you want to do any operations with this training image that has been added to a specific face identity.

There must be exactly one face visible per image. This REST function may take a few seconds to complete, because the system checks that exactly one face is clearly visible (otherwise, an error response is generated).

In the future, also image download URLs will be able to be used with the same easy add_face_image call syntax. Currently, only the valossaupload:// URLs created as a result of file uploads are supported.

Please make sure that each of the images is actually an image of the correct person. Typically, checking this involves some human work. Wrong images will deteriorate the quality of face detections.

Send an HTTP POST to the URL:

https://api.valossa.com/training/1.0/add_face_image

In this request, there is no need to specify the gallery, even when using a non-default gallery. The face ID is a unique identifier for the correct face.

Curl example:

curl --header "Content-Type:application/json" -X POST -d '{"api_key":"kkkkkkkkkkkkkkkkkkkk", "face":{"id":"bb254a82-08d6-4498-9ddb-3de4c88f1f66"}, "image":{"url":"valossaupload://ff357efe-1086-427d-b90c-1d1887fb1017"}}' https://api.valossa.com/training/1.0/add_face_image

Example response in an HTTP 200 OK message:

{
  "image_id": "8ac7ab90-44d1-4860-9a2f-2afbb175638a"
}

List existing face identities

Send an HTTP GET to the URL:

https://api.valossa.com/training/1.0/list_face_identities

Curl example if you are using your default gallery:

curl 'https://api.valossa.com/training/1.0/list_face_identities?api_key=kkkkkkkkkkkkkkkkkkkk'

Curl example if you are using a non-default gallery:

curl 'https://api.valossa.com/training/1.0/list_face_identities?api_key=kkkkkkkkkkkkkkkkkkkk&gallery_id=468a2b70-3b55-46f8-b209-8ad2fcabd5c8'

Example response in an HTTP 200 OK message:

{
  "face_identities": [
    {
      "id": "a99a59e3-ba33-4b00-8114-8bdd92a71dfa"
    },
    {
      "id": "bb254a82-08d6-4498-9ddb-3de4c88f1f66"
    }
  ]
}

It is also possible to list existing face identities with details.

Curl example if you are using your default gallery:

curl 'https://api.valossa.com/training/1.0/list_face_identities?api_key=kkkkkkkkkkkkkkkkkkkk&show_details=true'

Curl example if you are using a non-default gallery:

curl 'https://api.valossa.com/training/1.0/list_face_identities?api_key=kkkkkkkkkkkkkkkkkkkk&gallery_id=468a2b70-3b55-46f8-b209-8ad2fcabd5c8&show_details=true'

Example response in an HTTP 200 OK message:

{
  "face_identities": [
    {
      "id": "a99a59e3-ba33-4b00-8114-8bdd92a71dfa",
      "name": "Lizzy Blythriver",
      "gender": "female"
    },
    {
      "id": "bb254a82-08d6-4498-9ddb-3de4c88f1f66",
      "name": "Ricky Rickson",
      "gender": "male"
    }
  ]
}

List existing images added for a face identity

Send an HTTP GET to the URL:

https://api.valossa.com/training/1.0/list_face_images

In this request, there is no need to specify the gallery, even when using a non-default gallery. The face ID is a unique identifier for the correct face.

Curl example:

curl 'https://api.valossa.com/training/1.0/list_face_images?api_key=kkkkkkkkkkkkkkkkkkkk&face_id=bb254a82-08d6-4498-9ddb-3de4c88f1f66'

Example response in an HTTP 200 OK message:

{
  "face_images": [
    {
      "id": "8ac7ab90-44d1-4860-9a2f-2afbb175638a"
    },
    {
      "id": "b5559837-62a5-4f10-b250-a554ab2ce54c"
    }
  ]
}

Update face identity

Send an HTTP POST to the URL:

https://api.valossa.com/training/1.0/update_face_identity

In this request, there is no need to specify the gallery, even when using a non-default gallery. The face ID is a unique identifier for the correct face.

The "updates" structure contains one or more face parameters to update. The allowed parameters are "name" and "gender". The data type for these values is string. The maximum length of the value of "name" is 1024 characters. The value for "gender" is "male" or "female".

Note: To unset a field such as "name" or "gender" completely, just set it to null in an update_face_identity call. In an update, a value that is not mentioned in the "updates" structure will retain its old value if it had one (in other words, omitting the field from the update does not unset the value of the field, while setting it explicitly to null will unset it).

Curl example:

curl --header "Content-Type:application/json" -X POST -d '{"api_key":"kkkkkkkkkkkkkkkkkkkk", "face":{"id":"bb254a82-08d6-4498-9ddb-3de4c88f1f66", "updates":{"name":"Ricky Rixon-Borgmann"}}}' https://api.valossa.com/training/1.0/update_face_identity

Example response in an HTTP 200 OK message:

{}

Remove image from face identity

Send an HTTP POST to the URL:

https://api.valossa.com/training/1.0/remove_face_image

In this request, there is no need to specify the gallery, even when using a non-default gallery. The image ID is a unique identifier for the correct image.

Curl example:

curl --header "Content-Type:application/json" -X POST -d '{"api_key":"kkkkkkkkkkkkkkkkkkkk", "image":{"id":"b5559837-62a5-4f10-b250-a554ab2ce54c"}}' https://api.valossa.com/training/1.0/remove_face_image

Example response in an HTTP 200 OK message:

{}

Remove face identity

Send an HTTP POST to the URL:

https://api.valossa.com/training/1.0/remove_face_identity

In this request, there is no need to specify the gallery, even when using a non-default gallery. The face ID is a unique identifier for the correct face.

Curl example:

curl --header "Content-Type:application/json" -X POST -d '{"api_key":"kkkkkkkkkkkkkkkkkkkk", "face":{"id":"bb254a82-08d6-4498-9ddb-3de4c88f1f66"}}' https://api.valossa.com/training/1.0/remove_face_identity

Example response in an HTTP 200 OK message:

{}

Visualization of your results in a Valossa Report

Valossa Portal provides an easy-to-use visualization tool, called the Valossa Report, for you to get a quick visual overview of the most prominent detections, and also a more detailed heatmap for browsing the results. Remember to contact sales if you do not have access to Valossa Portal and Valossa Report yet.

On the home page, each displayed information box that is related to a successfully analyzed video contains a link to the Valossa Report of the video analysis results. To see examples of Valossa Report, click "Demos" on the home page (you must be logged in to Valossa Portal in order to do this).

Below you'll find example screenshots of Valossa Report.

(Actually the Valossa Report is a tool for viewing your Valossa Core metadata in an easy way for humans. When you're ready to integrate Valossa Core metadata to your application, please see the instructions for machine-reading the Valossa Core metadata.)

Overview

The Valossa Report's Overview gives you a quick visual overview of the analyzed video content.

The tags are an overview of the detected concepts. By clicking the arrows you can browse through the detections in the video. You can also search the concept within the video by clicking the magnifying glass symbol.

Heatmap

The Valossa Report's Heatmap displays the timeline of a video, and detections of concepts are placed on the timeline. Each detection is shown on its own row (its own timeline). Detections are grouped by their detection type such as human.face, visual.context, audio.context, etc. Please note that different colors are given to different detection types for distinguishing them visually.

Within a detection type, detections are grouped by prominence. For example, the most prominent faces are shown first.

With the Valossa Report controls, you can change the resolution of the timeline (how many distinct timeslots are shown) and the number of detections shown. You can also adjust the confidence threshold for several detection types. The detections below the chosen threshold are hidden.

The depth of color in the colored spots on the timeline for a detection shows how many detections of that concept are in that timeslot and/or how confident the detections are. Click on a colored spot, and the video player on the Valossa Report page will playback the video from the corresponding timecode. Thus, you are able to see the main concepts of the video arranged by time and prominence, and verify their correctness. With the main timeline and the seek bar under the video player, you can also move to any time-position in the video.

Tag & Train

The Tag & Train naming tool can be used to edit names and genders of the detected faces. Changes will be saved to the metadata of the video analysis job and indexed into the search automatically. Training functionality that allows the AI to learn from the changes is available.

Click the "Tag & Train" button above the face detections or the pencil next to a person name to open the tool.

General notes about metadata

Valossa Core metadata is provided in downloadable JSON files, which are available via the REST API (function job_results) or via the results page in Valossa Portal that shows the results and other info about your most recent jobs.

The sizes of the JSON files vary depending on the size of the videos and the number of detections, ranging from a few kilobytes to several megabytes. You should save the metadata JSON in your local database or file system. The metadata will not necessarily be stored perpetually in our system, download count limits may be imposed in the future, and it is also faster for your application to access the metadata from your local storage space.

The version number of the metadata format is continuously updated, when the format changes (version changelog of Valossa Core metadata). The version number is a concatenation of three integers, with a dot (.) as the delimiter: starting from the beginning of the string the version number x.y.z contains a major version number, a minor version number and a patch number. If only the patch version number (z in x.y.z) changes, the changes are purely additions to the structure i.e. they can't break the parsing code.

Output metadata JSON format

Basics: How (and why) to read Valossa Core metadata in your application code

Valossa Core metadata has been designed to address several needs. It answers questions such as:

What does the video contain?
Where — or when — are all the instances of X in the video?
What is in the video at any specific time position?
What is the video predominantly about?

Please see the images below for a quick explanation of how to read these things from the metadata.

Valossa Video Recognition AI addresses the needs 1 and 4 by detecting a varity of things and then ranking the most dominant detections from the video, so that the Valossa Core metadata can be used for answering questions such as "What are the visuals about?", "Who are the faces appearing in the video?", "What sounds are in the audio track?", "What are the spoken words about?", "What is the entire video about?", etc. The detections are grouped conveniently by the detection type, see more below. The needs 2 and 3 are addressed by Valossa Video Recognition AI with a smart time-coding logic that makes it easy to read either all the temporal occurrences of a specific detection or all the detections at a specific time position, whichever way is the most useful for your application.

A more detailed explanation of the fields "detections" and "by_detection_type" can be found in the subchapter Detections.

Detections are grouped by Valossa Video Recognition AI in a way that makes it easy for your application code to iterate over all instances (occurrences) of, for example, cats:

By reading the "by_second" field, your application code can easily list everything at a given time position. More details about the "by_second" field are provided in the subchapter Detections.

Using IAB categories, the metadata tells the topics of the video to your application code:

The main JSON structure

Valossa Core metadata about your videos is hierarchical and straightforward to parse for your application code. High-level structure of the current Valossa Core video metadata JSON format, not showing detailed subfields:

{
  "version_info": { "metadata_type": "core", "metadata_format": "...", "backend": "..." },
  "job_info": { "job_id": "...", "request": {...} },
  "media_info": { ... },
  "detections": { ... },
  "detection_groupings": { 
    "by_detection_property": { ... },
    "by_detection_type": { ... },
    "by_frequency": { ... },
    "by_second": [ ... ]
  },
  "segmentations": { ... }
}

Currently there are four supported values for the "metadata_type" field: "core", "frames_faces", "seconds_objects" and "frames_objects". The default type is "core" (Valossa Core metadata) — if you need "frames_faces" metadata that contains the bounding box information for the detected faces or "seconds_objects" or "frames_objects" metadata that contain the bounding box information for the detected visual objects, you must specify this in your API call when downloading metadata.

The version number of the metadata format (x.y.z, explained above) can be found in the "metadata_format" field under "version_info". Of course, the version numbering of Valossa Core metadata (the files with "core" as the value of the "metadata_type" field) is separate from the version numbering of "frames_faces" or "seconds_objects" or "frames_objects" metadata for the same video.

Using special metadata types such as "frames_objects" requires, obviously, that your video analysis job was run with capabilities that produce the special metadata files (in addition to the corresponding detection IDs in the Core metadata file).

You will best understand the details of the metadata structure by viewing an actual metadata JSON file generated from one of your own videos! As the very first thing you'll probably want to view your results using the easy Valossa Report visualization tool.

Note: In order to save storage space, JSON provided by the API does not contain line-breaks or indentations. If you need to view JSON data manually during your software development phase, you can use helper tools in order to get a more human-readable (pretty-printed) version of the JSON. For example, the JSONView plugin for your browser may be of help, if you download JSON metadata from the Portal: the browser plugin will display a pretty-printed, easily navigable version of the JSON. In command-line usage, you can use the "jq" tool or even Python: cat filename.json | python -m json.tool > prettyprinted.json

In the following subchapters, the JSON metadata format is described in more detail.

Detections

All concept detections from the video are listed in the field "detections". This is an associative array, where the key is a detection ID and the value is the corresponding detection. Please note that the detection ID is a string, and you must not assume that the string always represents an integer, even though the IDs often look like "1" or "237". So, the ID is a string, unique within the key space of the "detections" structure, but your code cannot assume that the string has a specific internal format.

The detection IDs are used in "detection_groupings" to refer to the specific detection, so the detailed information about each detection resides in one place in the JSON but may be referenced from multiple places using the ID. Inside the field "detection_groupings", four practical groupings of detections are given for you:

The subfield "by_detection_type" has detection type identifiers as the key and the value is an array of detection IDs; the array is sorted by relevance, most relevant detections first! Using "by_detection_type", you can easily for example list all the detected faces, or all the detected audio-based keywords. Want to find out whether there's a cat somewhere in your video? Just loop over the visual.context detections and match detections against Valossa concept identifier (cid) of "cat" ("02g28TYU3dMt"), against the Wikidata concept identifier of "cat" ("Q146"), against the Google Knowledge Graph concept identifier of "cat" ("/m/01yrx"), or even against the human-readable concept label "cat" if you're adventurous. See details below.
The subfield "by_second" contains an array, where each item corresponds to one second of the video. Using this array you can answer questions such as "What is detected at 00:45 in the video?". Under each second, there is an array of objects which contain at least the string-valued field "d" (detection ID). Using the detection ID as the index, you will find the detection from the "detections" list. If applicable, there is also the float-valued field "c" (confidence, max. 1.0), currently available only for visual.context and audio.context detections. If the field "o" exists, it contains an array of occurrence identifiers that correspond to this detection in this second.
The subfield "by_detection_property", introduced in Valossa Core metadata version 1.3.4, currently contains a convenient structure that helps you when several "human.face" detections have been matched to the same specific face ID in a face gallery. More information about iterating over faces based on the gallery face ID
The subfield "by_frequency", added in Valossa Core metadata version 1.3.11, summarizes recognized visual concepts into groups that describe a common "theme". This subfield contains lists of simultaneously appearing detections by their ids. The detections are listed in metadatas ["detections"] structure.

The following image helps understand the usage of detection IDs as references within the JSON data:

How to get an overview of the most prominent detections? That's easy: in "by_detection_type", start reading detections from the beginning of the lists under each detection type. Because the detections are sorted the most relevant ones ffirst, reading e.g. the 20 first detections from "human.face" gives you an overview of the most prominent faces in the video. For an easy and quick overview of detections, you may view the Valossa Report (visualization of detections) of the video in Valossa Portal.

However, please note that the "audio.speech" detections (speech-to-text results) and "audio.speech_detailed" detections (detailed speech-to-text results) are not ordered by prominence, as they are just raw snippets of speech detected from a specific time-range from within the video's audio track. The "audio.speech_detailed" detections, in their listing inside "by_detection_type", are ordered by the starting time of each word or utterance, so it's easy for you to iterate over the individual words (etc.) in the same order in which they can be heard on the video. The complete speech-to-text data of a video are also available in the SRT format from the Valossa Video Recognition API (see speech-to-text SRT download) and in Valossa Portal on the page that lists the most recent analysis results. The content of the downloadable SRT file is generated from the "audio.speech" detections from the Valossa Core metadata JSON file, so the information is the same whether you read the speech-to-text results from the metadata or from the SRT downloaded from Valossa Portal. Please note that the newlines in the generated speech-to-text SRT file are Unix-newlines (LF only, not CRLF).

Every detection in the JSON has, at minimum, the fields "t" (detection type identifier) and "label". The "label" is just the default human-readable label of the detected concept, and for many detection types, more specific information is available in additional data fields. The following is the list of currently supported detection type identifiers.

Fields that exist or don't exist in a detection, depending on the detection type and situation, include "occs", "a", "ext_refs", "categ" and "cid".

Detection types

The detection types are named in a logical, hierarchical way. The beginning of the name of each detection type tells something important about the detection:

visual.*: things detected from the visual content of the video frames
audio.*: things detected from the audio track of the video
human.*: specifically human-centered detections, currently comprising only the face-related detections
transcript.*: things detected from the pre-existing transcript, which was provided by you when you requested the analysis of the video
topic.*: topics of the video, detected either from the entire video (topic.iab and topic.general) or from time-based sections of the video (topic.iab.section)
explicit_content.*: deprecated types for explicit content — please note that many explicit visual detections are actually provided in visual.context detections, which are helpfully categorized using the Valossa tag categories in the metadata
external.*: keywords or phrases picked from something that is external to the video itself, currently supporting two information sources, which are the title of the video and the description of the video, both sources being provided by you in the new_job request

What is a "novelty word" (novelty_word) that appears in the names of various detection types? It means a keyword (or phrase) detected as relevant from spoken or textual content, and is so named because "novelty word" is an established term for keywords that represent the noteworthy or distinguishing features of the content. The term also underlines how novelty words are different from the other types of keywords which include, among other things, proper names such as names of locations or names of people.

Currently, the following detection types are supported.

visual.context
visual.object.localized
audio.context
audio.speech
audio.speech_detailed
human.face
human.face_group
transcript.keyword.compliance
transcript.keyword.novelty_word
transcript.keyword.name.person
transcript.keyword.name.location
transcript.keyword.name.organization
transcript.keyword.name.general
audio.keyword.compliance
audio.keyword.novelty_word
audio.keyword.name.person
audio.keyword.name.location
audio.keyword.name.organization
audio.keyword.name.general
external.keyword.novelty_word
external.keyword.name.person
external.keyword.name.location
external.keyword.name.organization
external.keyword.name.general
topic.iab
topic.iab.section
topic.iab.audio [DEPRECATED]
topic.iab.transcript [DEPRECATED]
topic.iab.visual [DEPRECATED]
topic.general
explicit_content.audio.offensive [DEPRECATED]
explicit_content.transcript.offensive [DEPRECATED]
visual.color
visual.text_region.full_frame_analysis
visual.text_region.lower_third
visual.text_region.middle_third
visual.text_region.upper_third

The identifiers are mostly self-explanatory. Please note that "visual.context" offers a broad range of visual detections such as objects; "audio.context" offers a broad range of audio-based detections; "topic.iab" and "topic.general" are categories for the entire video; "external.keyword.*" refers to keywords found from video description or title; "human.face_group" are people who have a temporal correlation high enough to probably have meaningful interaction with each other.

In addition to the clearly offensive, explicit words or phrases (such as swearwords) detected from speech and having the detection type "explicit_content.audio.offensive", any Content Compliance related keywords from speech or transcript have the detection types "audio.keyword.compliance" and "transcript.keyword.compliance", respectively. They include words and phrases related to violence, sex and substance use. Currently, these Content Compliance specific keyword detections are only available for English-language content.

Occurrences

The field "occs" contains the occurrence times of the detection. There is a start time and an end time for each occurrence. For example, a visual object "umbrella" might be detected 2 times: first occurrence from 0.3 seconds to 3.6 seconds, and another occurrence from 64.4 seconds to 68.2 seconds — so there would be 2 items in the "occs" array. Time values are given as seconds "ss" (seconds start) and "se" (seconds end), relative to the beginning of the video.

Detections that are not time-bound (such as topic.iab, topic.iab.* and external.keyword.*) cannot contain "occs".

If applicable to the detection type, occurrences have a maximum confidence ("c_max") detected during the occurrence period. (Because confidence varies at different moments during the occurrence, it makes sense to provide just the maximum value here. To find out the confidence during a particular moment, check out the "c" field of each second in the "by_second" data.) Currently, only visual.context and audio.context detections have "c_max".

Occurrences in a "topic.iab.section" detection, however, have a "c" field (not "c_max"). This is because the confidence value in them is for the entire occurrence instead of being a maximum picked from several confidences.

Please note that if you want to answer the question "What is in the video at time xx:xx?", then you should see the "by_second" array in the "detection_groupings". Occurrences, on the other hand, are good when you want to answer the question "At what time-sections is Y detected?"

Each occurrence also contains the ID of the shot where the occurrence starts. The shot ID, stored as an integer in the field "shs", is just a numerical index to the array "detected_shots" within segmentations (the first shot is at index 0, the next one is at index 1, and so on). Similarly, the field "she" provides the ID of the shot where the occurrence ends. These shot references make it easy to integrate detection occurrences to your video workflow that utilizes the shot boundaries for clipping or for a similar purpose. For example, when your use case involves finding Content Compliance related concepts (such as nudity-related concepts), the entire shot can be easily exported to your video editor application or MAM information system, instead of just the occurrence, if this is what your workflow needs.

Other optional data fields of a detection

As you remember, "t" and "label" are always given for a detection. The field "occs" might not be there. Besides "occs", there are other optional fields for a detection: "a", "ext_refs", "categ"

If exists, the object-field "a" contains attributes of the detection. For example, the "human.face" detections may have attributes: "gender" that includes the detected gender, "similar_to" that includes the possible visual similarity matches to persons in a face gallery, and "s_visible" i.e. the total screen-time of the face (note: nearly always less than the combined duration of the occurrences of the face, because during an occurrence some frames usually do not have the detected face or other concept — some frame-gaps are allowed in occurrences in order to create a practical simplification of time-bound visibility, while "s_visible" is the combined duration of only those frames where this face has actually been seen by the AI). The "gender" structure also contains the field "c" that provides the confidence of the detected gender (0.0 to 1.0).

If exists, the string-field "cid" contains the unique identifier of the concept in the Valossa Concept Ontology. All visual.context detections and audio.context detections have "cid". However, for example audio.speech detections don't have "cid".

If exists, the array-field "ext_refs" contains references to the detected concept in different ontologies. Most visual.context detections have "ext_refs", expressing the concept identity in an external ontology, such as the Wikiedata ontology or the Google Knowledge Graph ontology (or several ontologies, depending on the availability of the concept in the various external ontologies). Inside "ext_refs", the ontology identifier for Wikidata is "wikidata" and the ontology identifier for Google Knowledge Graph is "gkg" (see examples). If a specific external ontology reference object such as "wikidata" exists, there is an "id" field inside the object; the "id" field contains the unique identifier of the concept within that external ontology. Then you may search information about the concept from exteral services such as https://www.wikidata.org/. For "topic.iab" detections, the "ext_refs" field contains the ontology identifier "iab", and the ontology reference object describes the topic (IAB category) in the industry-standard IAB classification.

If exists, the object-field "categ" provides a useful list of the concept categories (such as "food_drink", "sport", "violence"...) for the detection. See detailed information on reading and understanding the detection categories.

Tag categories

If exists in a detection, the object-field "categ" contains the key "tags", and under the key "tags" there is an array-field that contains one or more category identifier tags (string-based identifiers such as "flora" or "fauna") for the concept of the detection. For example, a "dog" detection has the category tag array ["fauna", "pets"]. As another example, a "train station" detection has the category tag array ["place_scene", "public_transport", "traffic", "buildings_architecture"]. Many visual.context detections and some audio.context detections have "categ". Note! This is about the categories of a specific detection (a "single concept") — a completely different thing than the categories of the entire video (such as IAB categories).

By checking the "categ":"tags" of your detections, your reader code can easily filter detections, for example, to find all Content Compliance detections in your metadata. This is very useful in a Content Compliance use scenario, and Valossa Video Recognition AI is especially good at detecting Content Compliance related detections for the media industry (a typical use case is to detect audiovisual content that is inappropriate for a specific viewing time or viewer demographic). You may be interested in the blog post about Valossa reaching best accuracy in visual Content Compliance benchmark.

Currently, the following tag categories are supported.

Tag categories that are available through Valossa AI Core Models

Category name shown in Valossa Report	Identifier tag of the tag category in metadata	Explanation
Accidents and destruction	accident	Severe situations such as accidents, explosions, conflicts and destruction after natural or civil catastrophes.
Act of violence	act_of_violence	Act of violence that could injure a victim.
Automotive	automotive	Cars, trucks and motorcycles.
Aviation	aviation	Airplanes and spacecraft.
Bad language	bad_language	Bad language such as swearing.
Boats and ships	boats_ships	Boats and ships.
Bombs and explosions	bomb_explosion	Explosions and smokes.
Brand or product	brand_product	Brands and products such as branded vehicles. This category does not include logos. Logo detection model can be purchased separately.
Buildings and architecture	buildings_architecture	Different buildings and architectural details.
Celebrations and holidays	celebrations_holidays	Personal celebrations and holidays like wedding, graduation, Christmas and Thanksgiving.
Children, family and play	children_family_play	Childen, toys and accessories. Also board games and amusement park rides which can be enjoyed by people of all ages.
Computers and video games	computers_video_games	Computers and video games.
Consumer electronics	consumer_electronics	Mobile devices, gadgets, cameras, televisions, home appliances, etc.
Content compliance	content_compliance	This tag is present on all the Content Compliance concepts. In addition to this tag, a Content Compliance concept will also have a more specific tag. Usually, the more specific tag is one of these: act_of_violence, threat_of_violence, gun_weapon, injury, sensual, sexual, substance_use, violence, video_structure.
Crime	crime	Crime.
Death	death	Death.
Explicit content	explicit_content	Groups highly visual sexual and violent content. All concepts in this category also belong to more specific categories like "sexual", "violence" or "injury".
Fashion and wear	fashion_wear	Clothing, shoes, accessories, jewelry and makeup.
Animals	fauna	Animals.
Plants and mushrooms	flora	Plants.
Food or drink	food_drink	Foods and drinks, also eating and drinking.
Graphics	graphics	Graphics that enrich the media content.
Guns and weapons	gun_weapon	Weapons such as guns, knives and bows.
Hate speech	hate_speech	Hate speech.
Home and garden	home_garden	Interior design & decoration, furniture, rooms, home textiles, tableware, garden etc.
Basic human actions	human_basic	Basic human actions like sitting, standing, sleeping etc.
Human features and body parts	human_features	Human features and body parts, also some similar non-human body parts such as animal eyes. This category does not contain emotion or Content Compliance detections. Emotions (from faces) are available by contacting Valossa and requesting us to enable them. Content Compliance detections are in their specific categories.
Social situations and human life	human_life_social	Human activities in social contexts such as pride parade, gambling, stunt, dog walking, a student. Sports, professional and religional activities are not included, they can be found from separate categories.
Industrial	industrial	Machinery, power plants, cables etc.
Injury	injury	Signs of injury such as blood, wounds and bruises.
Lights and effects	lights_effects	Media enhancing light-based effects.
Materials	materials	Materials such as concrete, wood, iron etc.
Military equipment and people	military	Military staff, military vehicles, aircraft and vessels.
Military conflict	military_conflict	War, military conflict.
Music	music	Musical instruments, events and settings.
Natural disasters and severe weather	natural_disaster_severe_weather	Natural disasters such as flood and severe weather such as thunderstorm.
Natural phenomena	natural_phenomena	Natural phenomena and events.
Landscape and environment	nonlive_natural	Natural objects such as rock, mountain, sun, river and glacier.
Online piracy	online_piracy	Online piracy.
Other man-made objects	other_manmade_object	Objects, which are not in any other specific category.
Pets	pets	Dogs, cats and their accessories.
Place, location or scene	place_scene	Places and locations like living room, stadium or road.
Professions and work	professions_work	Humans at work. Please note that special professionals like athletes and military personnel are in their specific categories.
Public transport	public_transport	Public transport such as bus and train.
Religion	religion	Religion-related symbols, persons and places of worship.
Sensitive social issues	sensitive_social_issues	Sensitive social issues.
Sensual	sensual	Hinting towards sexuality such as bikinis and underwear, kissing, navel.
Sexual	sexual	Clearly sexual material and intimate body parts.
Sports	sport	Sports, sporting events and athletes. Please note that Valossa offers a separate football (soccer) event model, which is not included in the general sports category. We also would like to train a custom sports event model for your purposes (a specific type of sport etc.).
Sports equipment	sport_equipment	Sports equipment, protection and clothing.
Sport locations	sport_locations	Sport locations such as swimming pool.
Style	style	Image styles such as diagram or cartoon.
Substance use	substance_use	Smoking, drugs, medicines and alcohols.
Terrorism	terrorism	Terrorism.
Threat of violence	threat_of_violence	Threat towards a person, for example, aiming with a gun.
Traffic, traffic areas and signs	traffic	Traffic stations, parking lots and town squares, different roads, streets, paths, bridges and underpasses. Traffic lights, other signs and traffic congestion also belong to this category.
Travel destinations	travel_destinations	Travel destinations such as famous landmarks around the world.
Video structure	video_structure	Video structure elements such as black frame.
Violence, injuries and threats	violence	Acts and signs of violence.
Visual arts and crafts	visual_arts_crafts	Pieces and making of visual arts and crafts like sculptures, paintings and handicrafts.

Categories that are available through Extended Models (needs to be enabled separately by contacting sales)

Category name shown in Valossa Report	Category tag in metadata	Explanation
Football	football_soccer (19 concepts)	Any actions during the sport of soccer (a.k.a. football). This tag and the related detections are available only for those customers that have purchased the Valossa soccer (football) model separately.
Logo	logo (298 concepts)	Brand logos.

Find all Content Compliance related detections in an easy way

This is a practical example of using the "categ" information: how to find all those detections that are related to Content Compliance. You can loop over all the detections (or, if so desired, only all the detections having a particular detection type such as "visual.context") and check if the "categ":"tags" array contains the tag "content_compliance".

The same logic for reading "categ" entries works, of course, also for other tags than just "content_compliance". We are using "content_compliance" as the example, because Content Compliance is a popular use scenario for Valossa Video Recognition AI.

Example pseudo-code for finding all "content_compliance" tagged detections in the Core metadata of your video:

content_compliance_dets_by_det_id = []
foreach det_id --> det_item in metadata["detections"]:
    if exists det_item["categ"]:
        if det_item["categ"]["tags"] contains "content_compliance":
            content_compliance_dets_by_det_id[det_id] = det_item

Similar pseudo-code example but finding "content_compliance" tagged detections only from among the "visual.context" detections, not from among all detections of all types:

content_compliance_dets_by_det_id = []
if exists metadata["detection_groupings"]["by_detection_type"]["visual.context"]:
    foreach det_id in metadata["detection_groupings"]["by_detection_type"]["visual.context"]:
        det_item = metadata["detections"][det_id]:
        if exists det_item["categ"]:
            if det_item["categ"]["tags"] contains "content_compliance":
                content_compliance_dets_by_det_id[det_id] = det_item

Tips for reading some detection types

Speech-to-text or speech transcript

For "audio.speech" (speech-to-text) detections, the detected sentences/words are provided as a string in the "label" field of the detection. If you are one of those customers who have the "audio.speech_detailed" detections enabled, the individual words (or punctuation characters) provided by the "audio.speech_detailed" are similarly available in the "label" field of the detection. Of course, if you provided your own, pre-existing transcript for the video, there are no audio-based speech detections, but there are "transcript.*" detections with the content in the "label" fields.

If you provided your own, pre-existing transcript of the speech, the words and phrases picked from that are available in the "transcript.*" detections, not "audio.speech" or "audio.speech_detailed".

Voice emotion

If your usage configuration includes voice emotion (emotion audible from voice prosodics, not from words or phrases themselves) as part of the analysis, the voice emotion information is provided in the "audio.voice_emotion"-referencing data items in the by_second grouping structure. See the more detailed instructions in the section on emotion-related results from faces and speech.

IAB advertisement categories for entire videos or for time-based sections inside a video

Information related to "topic.iab" and "topic.iab.section" detections: By matching the standardized IAB advertisement categories of your videos to your ads inventory, and you'll be able to display mathing ads when displaying a specific video. Read the "topic.iab" detections to get the IAB categories of the entire video, in other words, to make decisions for the entire video file. Read the "topic.iab.section" detections to get the IAB categories that have been detected for some time-based sub-sections of the video. Unlike "topic.iab" detections, the "topic.iab.section" detections contain occurrences ("occs"), because they are time-bound detections and so it makes sense to have occurrences (containing most importantly the start time and end time) for them. Please also note that the "topic.general" detections are separate from the "topic.iab.*" detections. The "topic.general" detections could be used for ads placement in some cases, because they aim to describe the topics of the video, but the IAB categories are obviously more directly advertisement-oriented. If IAB categories have the IAB taxonomy version 2.1 (the taxonomy version be seen in the items in the metadata), the "ext_refs" structure of each such category will a boolean-valued "is_scd_category" field, which is boolean true if the category is SCD (Special Category Data) related to sensitive matters (to protect consumer privacy in the context of advertising), otherwise boolean false.

OCR (visual recognition of text and numbers)

Information related to "visual.text_region.full_frame_analysis", "visual.text_region.lower_third", "visual.text_region.middle_third" and "visual.text_region.upper_third" detections: These OCR detection types are named based on the rough location of the text on the video frames. This allows, for example, the recognition of subtitles, which are usually placed in the lower third of videos (detection type: "visual.text_region.lower_third"). Of course, the "visual.text_region.full_frame_analysis" detections contain results from looking at the entire frame. Occurrences (occs) in OCR detections correspond to time-sections where the AI detected the particular text region in a continuous manner. In the future, there will also be separate detection types for keywords, phrases etc. that are picked from the OCR content. An example of OCR results is provided in the detection data examples.

Face gallery matches and face sizes

Information related to "human.face" detections: If and only if a face is similar to one or more faces in a face gallery (either the Valossa Celebrities Gallery or your custom face gallery), the "a" field of the specific face detection object will contain a "similar_to" field that contains an array of the closely matching faces (from a gallery), and within each item in the "similar_to" array there is a string-valued "name" field providing the name of the visually similar person and "c" that is the float-valued confidence (0.0 to 1.0) that the face is actually the named person. In "similar_to", the matches are sorted the best first. Please note that a face occurrence doesn't directly contain "c" — confidence for faces is only available in the "similar_to" items. Starting from Valossa Core metadata version 1.3.4, each "similar_to" item contains a "gallery" field and a "gallery_face" field. The "gallery" field contains an "id" field, the value of which is the face gallery ID (a UUID) of the gallery from which the matched face identity was found (see custom galleries). The "gallery_face" field contains a string-valued "name" field and an "id" field, the value of which is the face ID (a UUID), that is, a unique identifier of the face identity (person) within the specified face gallery. (Note: A "name" field exists in two places, for reader code compatibility with previous metadata formats.) For information about how to get the face coordinates (bounding boxes) of a face and how to find all occurrences of a specific gallery-matched face identity, see the separate subsection.

Information related to second-based timeslots (contained in the by_second grouping structure) for "human.face" detections: Starting from Valossa Core metadata version 1.3.12, information about the size of each face is provided per second. For each second-based timeslot of a "human.face" detection, the "a" (attributes) field contains the "sz" (size) field, which contains the "h" (height) field. The "h" field is float-valued and the value is given as relative to the height of the video frames, in other words, the value 1.0 means the full height of the video frame. While a face is typically seen in multiple frames within a particular second, the value of "h" is simply the height of this face at the moment when it is first encountered within the second. If you need complete, frame-by-frame size information (width, height, position for each face in each frame) for "human.face" detections, please see face bounding boxes in frames_faces metadata.

Example of face height data at a particular second for a particular face:

...
{
  "d": "1",
  "o": ["1"],
  "a": {
      "sz": {"h": 0.188}
  }
},
...

Explicit content (for content moderation cases) in visual.context detections

The old detection type "explicit_content.nudity" has been deprecated and is not available in any jobs created after 2018-12-05. Instead of that deprecated detection type, you should use the new, more detailed, more accurate explicit content model. (Explicit content is, for example, nudity-related, violence-related or substance-use-related content.) The new model has been integrated as part of the "visual.context" detections. Those "visual.context" detections, which are related to explicit content, have such category tags that they can be easily distinguished from the non-explicit detections. See more information on detection category tags

Colors detected on each second

The detected dominant colors in the video, per second, are provided in such a way that there is only one "visual.color" detection, which covers the entire video. The colors are provided as RGB values (6 hexadecimal digits in a string) stored as attributes in the "by_second" structure in those second-based data items, where the "d" field refers to the single detection that has the detection type "visual.color". Please note that at the end of a video, there might be a second where the color item is not available, so please do not write reader code that assumes a "visual.color" detection reference to exist at absolutely every second of the video. The attributes are in the "a" field of the color-related data item of the given second, and the "a" field contains an array-valued "rgb" field, where each item is an object containing information about a particular detected color. In each of those objects, the "f" field is the float-valued fraction (max. 1.0) of the image area that contains the particular color (or contains a close-enough approximation of that color), and the "v" field is the RGB value. The "letter digits" (a-f) in the hexadecimal values are in lowercase.

Example of color data at a particular second:

...
{
  "d": "12",
  "o": [
    "13"
  ],
  "a": {
    "rgb": [
      {
        "f": 0.324,
        "v": "112c58"
      },
      {
        "f": 0.301,
        "v": "475676"
      },
      {
        "f": 0.119,
        "v": "9f99a3"
      }
    ]
  }
},
...

Detection data examples

An example of what is found in "detections", a visual.context detection:

...
"86": {
  "t": "visual.context",
  "label": "hair",
  "cid": "lC4vVLdd5huQ",
  "ext_refs": {
    "wikidata": {
      "id": "Q28472"
    },
    "gkg": {
      "id": "/m/03q69"
    }
  },
  "categ": {
    "tags": [
      "human"
    ]
  },
  "occs": [
    {
      "id": "267",
      "ss": 60.227,
      "se": 66.191,
      "c_max": 0.80443,
      "shs": 47,
      "she": 48
    },
    {
      "id": "268",
      "ss": 163.038,
      "se": 166.166,
      "c_max": 0.72411,
      "shs": 131,
      "she": 131
    }
  ]
},
...

Another example from "detections", a human.face detection:

...
"64": {
  "t": "human.face",
  "label": "face",
  "a": {
    "gender": {
      "c": 0.929,
      "value": "female"
    },
    "s_visible": 4.4,
    "similar_to": [
      {
        "c": 0.92775,
        "name": "Tina Schlummeister"
        "gallery": {
          "id": "a3ead7b4-8e84-43ac-9e6b-d1727b05f189"
        },
        "gallery_face": {
          "id": "f6a728c6-5991-47da-9c17-b5302bfd0aff",
          "name": "Tina Schlummeister"
        }
      }
    ]
  },
  "occs": [
    {
      "id": "123",
      "ss": 28.333,
      "se": 33.567,
      "shs": 8,
      "she": 9
    }
  ]
},
...

An example of an audio.context detection:

...
"12": {
  "t": "audio.context",
  "label": "exciting music",
  "cid": "o7WLKO1GuL5r"
  "ext_refs": {
    "gkg": {
      "id": "/t/dd00035"
    }
  },
  "occs": [
    {
      "id": "8",
      "ss": 15.0,
      "se": 49.0,
      "shs": 14,
      "she": 29,
      "c_max": 0.979
    }
  ]
},
...

An example of two audio.speech_detailed detections, one containing a word (with a confidence value and with diarization, that is, speaker identification in the "s" structure of the attributes of the detection) and one containing a punctuation character, which, obviously, has a zero duration, because a punctuation character does not consume actual time during the detected speech:

...
"210": {
  "t": "audio.speech_detailed",
  "label": "stay",
  "c": 0.59,
  "a": {"s": {"id": "14"}},
  "occs": [
    {
      "id": "341",
      "ss": 44.32,
      "se": 44.4,
      "shs": 28,
      "she": 28
    }
  ]
},
...
"318": {
  "t": "audio.speech_detailed",
  "label": ".",
  "c": 1.0,
  "a": {"s": {"id": "6"}},
  "occs": [
    {
      "id": "449",
      "ss": 100.6,
      "se": 100.6,
      "shs": 77,
      "she": 77
    }
  ]
},
...

If you are not interested in individual words or their diarizations, you may want to use the "audio.speech" detections instead of "audio.speech_detailed" detections. The "audio.speech" detections contain the same speech but concatenated into chunks that roughly correspond to typical subtitle groupings (without diarization, without word-level start-and-end timestamps, without confidence values).

Please note that the "audio.speech_detailed" detections may not be available to all customers using the speech-to-text capabilities of Valossa AI. Those customers with speech-to-text but without the "audio.speech_detailed" support will only have the "audio.speech" detections in the metadata (however, the AI-picked "audio.keyword.*" detections are available for both kinds of speech-to-text customers).

An example of a visual.text_region.* (OCR, text region) detection, in this case specifically a visual.text_region.lower_third detection that happens to contain two lines (rows) of subtitles recognized in the lower-third area of the video frames, along with line-specific confidence values:

...
"26": {
  "t": "visual.text_region.lower_third",
  "label": "text region",
  "a": {
    "lang": "en",
    "text": {
      "lines": [
        "and the winner of the competition",
        "is from the outskirts of the capital"
      ],
      "lines_with_splitting": [
        ["and", "the", "winner", "of", "the", "competition"],
        ["is", "from", "the", "outskirts", "of", "the", "capital"]
      ],
      "as_one_string": "and the winner of the competition is from the outskirts of the capital",
      "c": {
        "lines": [
          1.0,
          0.921
        ]
      }
    }
  },
  "occs": [
     {
      "id": "27",
      "ss": 11.3,
      "se": 15.02,
      "shs": 4,
      "she": 5
     }
   ]
},
...

An example of an IAB category detection:

...
"173": {
  "t": "topic.iab",
  "label": "Personal Finance",
  "ext_refs": {
    "iab": {
      "labels_hierarchy": [
        "Personal Finance"
      ],
      "id": "IAB13"
    }
  }
},
...

An example of keyword detection:

...
"132": {
  "t": "transcript.keyword.name.location",
  "label": "Chillsbury Hills",
  "occs": [
    {
      "ss": 109.075,
      "se": 110.975,
      "id": "460"
    }
  ]
}
...

Please note that transcript keyword occurrence timestamps are based on the input SRT timestamps. In the future, if a non-timecoded transcript is supported, transcript keywords might not have occurrences/timecoding.

How to find all face occurrences of a recognized person, and how to read the face coordinates (bounding boxes)

When viewing your video analysis results, you may have noticed that several different "human.face" detections (under different detection IDs) may be recognized as the same named person from a face gallery. This is natural, because to the AI, some face detections seem different enough from each other so they are classified as separate faces (separate face detections)... but each of those detections is similar enough to a specific face in the gallery, so each of the detections has a "similar_to" item for the same gallery face! For example, there could be two "human.face" detections which are "similar_to" the gallery face "Steve Jobs", the other one with confidence 0.61 and the other one with confidence 0.98.

Of course, a question arises: Is there an easy way to see list all the detections of a specific gallery face within a given Valossa Core metadata file? For example, find all "human.face" detections that were (with some confidence) matched to the gallery face "Steve Jobs"? Yes, there is.

Under "detection_groupings":"by_detection_property", certain types of detections are grouped by their certain shared properties. Currently, the only supported property-based grouping is for "human.face" detections, and for them the only supported property-based grouping has the identifier "similar_to_face_id". As shown in the example below, all detected faces that have at least one "similar_to" item (with a gallery face ID) are listed in the structure, indexed by the gallery face ID. A gallery face ID is a UUID that uniquely identifies the specific person (more precisely: the specific face identity) within the particular face gallery.

Please note that some legacy gallery faces might not have a face ID (UUID) and thus cannot be found in the "similar_to_face_id" structure. This restriction only applies to a few customers who have a face gallery that was created before the introduction of the "similar_to_face_id" grouping structure, and of course to the analysis jobs that have been run using an old version of the system: the face IDs and the "similar_to_face_id" grouping structure were introduced in the version 1.3.4 of Valossa Core metadata.

Under each gallery face ID, there is an object that contains the fields "moccs" and "det_ids".

In the "moccs" field, there is an array of objects that are the merged occurrences of the one or more "human.face" detections that share a specific gallery face ID in their "similar_to" items. The naming "moccs" highlights the difference of the format to the "occs" format that can be found in the actual "human.face" detections.

In the "det_ids" field, there is an array of the IDs of the detections that have this specific gallery face ID in their "similar_to" items. Thus, if you want to read all the original corresponding "human.face" detections (including, among other things, the original occurrences separately for each detection in a "non-merged form") for any specific gallery face ID, it is easy.

Of course, if there is only one "human.face" detection having a "similar_to" item with a given gallery face ID, then there is only one detection ID in the "det_ids" array under that gallery face ID, and the "moccs" array of that gallery face originates solely from the occurrences of the single corresponding "human.face" detection.

The name of each face is available in the "similar_to" items of the "human.face" detections, which are referred to with their detection IDs listed in "det_ids". So, for example, by looking at the item at index "3" in the "detections" field of the metadata you would see that the face ID "cb6f580b-fa3f-4ed4-94b6-ec88c6267143" is "Steve Jobs". Naturally, an easy way for viewing the "merged" faces information is provided by the Valossa Report tool.

Example of "similar_to_face_id" detection groupings data, where the occurrences of the face detections "3" and "4" with similarity to Steve Jobs (cb6f580b-fa3f-4ed4-94b6-ec88c6267143) have been merged into one easy-to-parse "moccs" structure:

...
"detection_groupings": {
  ...
  "by_detection_property": {
    "human.face": {
      "similar_to_face_id": {
        "cb6f580b-fa3f-4ed4-94b6-ec88c6267143": {
          "moccs": [
            {"ss": 5.0, "se": 10.0},
            {"ss": 21.0, "se": 35.0},
            {"ss": 64.0, "se": 88.0},
            {"ss": 93.0, "se": 98.0},
            {"ss": 107.0, "se": 112.0},
            {"ss": 123.0, "se": 137.0},
            {"ss": 157.0, "se": 160.0},
            {"ss": 196.0, "se": 203.0},
            {"ss": 207.0, "se": 212.0}
          ],
          "det_ids": ["3", "4"]
        },
        "648ec86d-4d91-42a6-928d-a25d8dc2691c": {
          "moccs": [
            {"ss": 194.0, "se": 197.0},
            {"ss": 229.0, "se": 237.0}
          ],
          "det_ids": ["19"]
        },
        ...
      }
    }
  },
  ...
},
...

Do you need face coordinates, that is, the bounding boxes for each detected face at a specific point in time? They are available from the Valossa Video Recognition API, but because of the considerable file size, the bounding boxes are not part of the Valossa Core metadata JSON. The face bounding box data must be downloaded as a separate JSON file from the API. The metadata type identifier of this special metadata JSON is "frames_faces" (in "version_info":"metadata_type"). When downloading the metadata, you need to specify the metadata type with the parameter "type=frames_faces" in the job_results call.

Please note that "frames_faces" metadata may not be available for your old video analysis jobs, as the feature has not always been part of the system.

The "frames_faces" metadata is easy to parse. The "faces_by_frame" field, which always exists, is an array that is indexed with the frame number so that the information for the first frame is at [0], the information for the next frame at [1] and so on. For each frame, there is an array that contains one bounding box object for each face that was detected in that frame. Of course, a frame without any detected faces is represented by an empty array.

Every bounding box object contains the fields "id", "x", "y", "w", "h". The value of "id" is the same detection ID that the corresponding "human.face" detection has in the Valossa Core metadata file of the same video analysis job. The values of "x" and "y" are the coordinates of the upper-left corner of the bounding box (the x offset from the left edge of the frame, and the y offset from the top of the frame). The values of "w" and "h" are the width and height of the bounding box, respectively. The values of "x", "y", "w", "h" are all given as float values relative to frame size, thus ranging from 0.0 to 1.0, with the following exception. Because a detected face can be partially outside the frame area, some face coordinates may be slightly less than 0.0 or more than 1.0 in the cases where the system approximates the edge of the invisible part of a bounding box. For example, the "x" coordinate of a face in such a case could be -0.027.

Example job_results request for "frames_faces" metadata, using HTTP GET:

https://api.valossa.com/core/1.0/job_results?api_key=kkkkkkkkkkkkkkkkkkkk&job_id=167d6a67-fb99-438c-a44c-c22c98229b93&type=frames_faces

Example response in an HTTP 200 OK message:

{
  "version_info": { "metadata_type": "frames_faces", "metadata_format": "...", "backend": "..." },
  "job_info": { ... },
  "media_info": { ... },
  "faces_by_frame": [
    ...
    [],
    [],
    [],
    [],
    [
      {
        "id": "1",
        "x": 0.4453125,
        "y": 0.1944444477558136,
        "w": 0.11953125149011612,
        "h": 0.21388888359069824
      }
    ],
    [
      {
        "id": "1",
        "x": 0.4351562559604645,
        "y": 0.19583334028720856,
        "w": 0.11953125149011612,
        "h": 0.2152777761220932
      }
    ],
    [
      {
        "id": "1",
        "x": 0.42578125,
        "y": 0.19722221791744232,
        "w": 0.12187500298023224,
        "h": 0.22083333134651184
      },
      {
        "id": "5",
        "x": 0.3382812440395355,
        "y": 0.23888888955116272,
        "w": 0.20468750596046448,
        "h": 0.3986110985279083
      }
    ],
    ...
  ]
}

Localized objects (visual objects with bounding boxes)

In addition to the coordinates for faces, the Valossa Video Recognition API also provides coordinates ("bounding boxes") for certain types of visual detections. Currently, however, this feature is available only for logos.

The relevant detection type is "visual.object.localized". Thus, in order to know which visual object detections have coordinates, your code that processes your Valossa Core metadata needs to read the "visual.object.localized" detections (not "visual.context").

Much like the "frames_faces" metadata containing the bounding boxes for "human.face" detections, the bounding box coordinates for the "visual.object.localized" detections are stored in a separate JSON file. This separate JSON file contains the "seconds_objects" metadata and can be downloaded using the API by specifying the metadata type "seconds_objects" in the job_results request. Because all detections are listed by their IDs in the Valossa Core metadata file (the main metadata file), also the "visual.object.localized" detections can be found in the "detections" structure within the Valossa Core metadata, but the time-specific, changing coordinates of their bounding boxes must be read from the "seconds_objects" metadata.

Please note that "seconds_objects" metadata may not be available for your old video analysis jobs, as the feature has not always been part of the system.

Again similar to the "frames_faces" metadata, the coordinates in "seconds_objects" metadata are given as float values relative to the picture frame size. The values range from 0.0 to 1.0, except that in some corner cases the coordinates may be slightly less than 0.0 or more than 1.0, so please take these possibilities into account in your reader code. Every bounding box object contains the fields "x", "y", "w", "h", "c". The values of "x" and "y" are the coordinates of the upper-left corner of the bounding box (the x offset from the left edge of the frame, and the y offset from the top of the frame). The values of "w" and "h" are the width and height of the bounding box, respectively. The value of "c" is the confidence of the detection in that bounding box area during that second.

Whereas the "frames_faces" metadata contains faces for each frame, the "seconds_objects" metadata contains objects for each second (not frame). The "objects_by_second" array always exists in the "seconds_objects" metadata. The "objects_by_second" array is indexed with the second number so that the information for the first second is at [0], the information for the next second at [1] and so on. Each second in the "objects_by_second" array is represented by an array, which of course may be empty, if there are no detected objects with bounding boxes in that specific second. In the array of a specific second, the items are detection-specific; there might be several different detections on the same second, for example, a Visa logo and a MasterCard logo. The detection ID is in the field "d" within the detection-specific item, and naturally details of the corresponding concept can be found under that specific detection ID in the "detections" structure of the Core metadata of the same video. The "seconds_objects" metadata does not contain any information on the actual detected concepts, because all that information is already available in the single correct place: the Core metadata of the video.

In the "seconds_objects" metadata, the occurrence(s) overlapping with the particular second-based timeslot of the particular "visual.object.localized" detection are referred to with their occurrence IDs in the array-valued "o" field, should you ever need to find the corresponding occurrences of a detection in an easy way. For your convenience, the detection and occurrence reference mechanism — based on the familiar "d" and "o" fields — has been deliberately designed to be similar to the mechanism used in the generic "by_second" structure of Valossa Core metadata. And yes, the "visual.object.localized" detections have references to them also in the "by_second" structure in the Core metadata, being consistent with the Core metadata specification, but the bounding box coordinates are only available in the separate "seconds_objects" metadata JSON, which exists to limit the size of the Core metadata file.

The bounding box coordinates for a specific detection are provided in the array-valued "b" field. For each "visual.object.localized" detection (e.g. a Coca-Cola logo), there might be one or more bounding boxes during a given second; in other words, the "b" array could contain more than one bounding box. For example, if there are two Coca-Cola logos simultanenously in the picture and the detection ID of the Coca-Cola logo "visual.object.localized" detection happens to be "137" in this particular video, then there are two bounding box items for the detection ID "137". Please note that in the "by_second" section in your Core metadata, the confidence value "c" of a "visual.object.localized" detection is the highest confidence of the possibly multiple simultaneously observed bounding boxes of the same detection (e.g. two simultanenous images of the same logo) during that one-second-long timeslot; to see the confidences (and coordinates) of the possibly multiple bounding boxes, your reader code needs to examine the "seconds_objects" metadata.

Please also note the existence of the frames-based "frames_objects" metadata for "visual.object.localized" detections. This metadata type has been introduced more recently than "seconds_objects" metadata: the metadata type "frames_objects" was made available in 2020.

Example job_results request for "seconds_objects" metadata, using HTTP GET:

https://api.valossa.com/core/1.0/job_results?api_key=kkkkkkkkkkkkkkkkkkkk&job_id=167d6a67-fb99-438c-a44c-c22c98229b93&type=seconds_objects

Example response in an HTTP 200 OK message:

{
  "version_info": { "metadata_type": "seconds_objects", "metadata_format": "...", "backend": "..." },
  "job_info": { ... },
  "media_info": { ... },
  "objects_by_second": [
    ...
    [],
    [],
    [
      {
        "b": [
          {
            "x": 0.2671875,
            "y": 0.7472222222222222,
            "w": 0.08072916666666667,
            "h": 0.06574074074074074,
            "c": 0.995
          },
          {
            "x": 0.4979166666666667,
            "y": 0.6685185185185185,
            "w": 0.08072916666666667,
            "h": 0.06944444444444445,
            "c": 0.984
          }
        ],
        "d": "137",
        "o": ["314"]
      }
    ],
    [
      {
        "b": [
          {
            "x": 0.49583333333333335,
            "y": 0.6685185185185185,
            "w": 0.08489583333333334,
            "h": 0.06851851851851852,
            "c": 0.991
          },
          {
            "x": 0.2713541666666667,
            "y": 0.7435185185185185,
            "w": 0.07760416666666667,
            "h": 0.06759259259259259,
            "c": 0.865
          }
        ],
        "d": "137",
        "o": ["314"]
      },
      {
        "b": [
          {
            "x": 0.9083333333333335,
            "y": 0.23333333333333,
            "w": 0.0101563746437,
            "h": 0.467467467467467,
            "c": 0.991
          }
        ],
        "d": "138",
        "o": ["315"]
      }
    ],
    [],
    ...
  ]
}

Speech-to-text results

The speech-to-text results of the video analysis are available in the SRT format from the Valossa Video Recognition API. The same SRT file can also be manually downloaded in Valossa Portal.

When downloading the speech-to-text SRT, you need to use the parameter "type=speech_to_text_srt" in the job_results call. The language of the speech-to-text operation is the one that was specified in the language selection of the new_job call that created the video analysis job.

Please note that the newlines in the generated speech-to-text SRT file are Unix-newlines (LF only, not CRLF). The same speech-to-text information is also available in the "audio.speech" detections of the Core metadata of the video. The SRT is provided for your convenience; the SRT file makes it easier to achieve interoperability with the various systems that expect time-coded speech information in the SRT format.

Example job_results request for speech-to-text results in the SRT format, using HTTP GET:

https://api.valossa.com/core/1.0/job_results?api_key=kkkkkkkkkkkkkkkkkkkk&job_id=167d6a67-fb99-438c-a44c-c22c98229b93&type=speech_to_text_srt

Example response in an HTTP 200 OK message:

1
00:00:01,910 --> 00:00:04,420
hello James and Jolie

2
00:00:04,420 --> 00:00:08,120
you shouldn't go there you know

3
00:00:08,119 --> 00:00:13,639
I had no idea that this could work
please listen to me this is so fabulous

Sentiment (emotion, expression) results from faces and speech

These analysis results are not available unless face & speech emotion analytics is separately activated for you before analyzing your videos. Contact sales to request activation of the feature.

There are three different kinds of sentiment and emotion related information in the metadata. How your software can read each of them is described in the following.

Valence from faces:

Valence is a form of sentiment that describes the emotional positivity or negativity of a person at a specific moment in time.

In the "by_second" structure, every "human.face"-related seconds-based item, if its valence has been detected by the AI, has a "sen" structure for the sentiment (inside an "a" field for attributes). In the "sen" structure, you will find a float-valued "val" field providing the valence of the specific face on the specific 1-second time interval. Valence ranges from -1.0 (most negative) to 1.0 (most positive), 0.0 being a neutral valence.

Example:

{
...
  "a": {"sen": {"val": -0.82}, ...},
  "d": "9",
  "o": ["51"]
...
}

Named emotions (expressions) from faces:

Several emotional states can be recognized on faces. What emotions are supported? Please see the following explanation:

Most customers have the new V2 face expressions in use. In V2, the named emotions are: joy, mild joy, sadness, serious expression, fear, tension/anxiousness, disgust, displeasure, anger, concentration/displeasure, surprise, startlement, neutral
Some long-time customers (pre December 2020), because of system backward-compatibility reasons, may have the old V1 face expressions in use. In V1, the named emotions are: happiness, sadness, anger, disgust, surprise, neutral

Named emotions, with confidences (max. 1.0), are provided for a face at a specific moment in time. The identifier strings for the emotions are the same as the emotion names listed above (please note the V2 vs. V1 distinction regarding the available emotion identifiers).

In the "by_second" structure, every "human.face"-related seconds-based item, if its emotional state has been detected by the AI, has a "sen" structure for the sentiment (inside an "a" field for attributes). In the "sen" structure, you will find an "emo" field providing the emotions of the specific face on the specific 1-second time interval. The field "emo" is an array, because sometimes more than one emotion can be detected from the same face. In each item in the array, you will find a "value" field providing the emotion identifier string and a float-valued "c" field providing the confidence (maximum 1.0).

Example, showing also valence in addition to a named emotion for this specific face during the specific 1-second interval:

{
...
  "a": {"sen": {"emo": [{"c": 0.772, "value": "disgust"}], "val": -0.796}}, 
  "d": "1",
  "o": ["1"]
...
}

Valence from speech:

Speech-based sentiment is currently available for English only. It contains valence information, which describes the emotional positivity or negativity of speech fragments that are heard on the video.

In the "detections" structure, each "audio.speech" detection has a "sen" structure for sentiment (inside an "a" field, for the attributes of the detection) if its valence has been detected by the AI. In the "sen" structure, you will find the valence of the speech fragment, in a float-valued "val" field. Valence ranges from -1.0 (most negative) to 1.0 (most positive), 0.0 being a neutral valence.

Example:

{
...
  "t": "audio.speech",
  "label": "we profoundly believe that justice will win despite the looming challenges",
  "a": {"sen": {"val": 0.307}, ...}
... 
}

Emotion from voice:

Please note that voice emotion is NOT the same thing as speech sentiment. Voice emotion is the emotional state detected from the sounds of human voices, while speech sentiment is the sentiment derived from the actual content (sentences, phrases, words) of the speech.

If your usage configuration includes voice emotion as part of the analysis, the voice emotion information is found in the seconds-based data of an "audio.voice_emotion" detection. (There was previously an alternative format for voice emotion data, but that format has been deprecated in favor of the new, more flexible format.)

The detected voice emotions in the video, per second, are provided in such a way that there is only one "audio.voice_emotion" detection, which covers all voice emotions detected at any time-point in the video. The voice emotions are provided as attributes (inside an "a" field) in the "by_second" structure in those second-based data items, where the "d" field refers to the single detection that has the detection type "audio.voice_emotion". Such an "a" field contains two fields: a float-valued "val" field, which provides the detected emotional valence, ranging from -1.0 to 1.0, and a float-valued "aro" field, which provides the detected emotional arousal, ranging from 0.0 (minimum arousal) to 1.0 (maximum arousal). Naturally, there might be several occurrences (occs) within that one "audio.voice_emotion" detection, if the video contains several discontiguous sections of detected voice-emotional material.

Example of voice emotion data at a particular second:

{
...
  "a": {"val": -0.022, "aro": 0.655},
  "d": "1034",
  "o": ["1688"]
...
}

Segmentations

In "segmentations", the video is divided into time-based segments using different segmentation rules.

Currently we support automatically detected shot boundaries, hence "segmentations" contains "detected_shots". The array "detected_shots" in segmentations provides shot boundaries, as an object for each detected shot, with seconds-based start and end timepoints (float-valued fields "ss", "se") and with start and end frame numbers (integer-valued fields "fs", "fe"). The shot duration as seconds is also provided (float-valued field "sdur"). Note: frame-numbers are 0-based, i.e. the first frame in the video has the number 0. All the fields "ss", "se", "fs", "fe", "sdur" are found in every shot object. The ordering of the shot objects in the array "detected_shots" is the same as the ordering of the detected shots in the video.

Example data:

"segmentations": {
  "detected_shots": [
    {
      "ss": 0.083,
      "se": 5.214,
      "fs": 0,
      "fe": 122,
      "sdur": 5.131
    },
    {
      "ss": 5.214,
      "se": 10.177,
      "fs": 123,
      "fe": 241,
      "sdur": 4.963
    },
    ...
  ]
}

Shots are referred to in other parts of the metadata: Occurrences (within concept detections) contain the 0-based index of the shot, during which the occurrence starts. This makes your video workflow integration easier in cases, where the beginning of the ongoing shot is important in relation to the detected concept.

Code examples for reading metadata

Example code snippet (in Python) that illustrates how to access the data fields in Valossa Core metadata JSON:

import json
metadata = None
with open("your_core_metadata.json", "r") as jsonfile:
        metadata = json.loads(jsonfile.read())

# Loop over all detections so that they are grouped by the type
for detection_type, detections_of_this_type in metadata["detection_groupings"]["by_detection_type"].iteritems():
        print "----------"
        print "Detections of the type: " + detection_type + ", most relevant detections first:"
        print
        for det_id in detections_of_this_type:
                print "Detection ID: " + det_id
                detection = metadata["detections"][det_id]
                print "Label: " + detection["label"]
                print "Detection, full info:"
                print detection

                # Example of accessing attributes (they are detection type specific)
                if detection_type == "human.face":
                        attrs = detection["a"]
                        print "Gender is " + attrs["gender"]["value"] + " with confidence " + str(attrs["gender"]["c"])
                        if "similar_to" in attrs:
                                for similar in attrs["similar_to"]:
                                        print "Face similar to person " + similar["name"] + " with confidence " + str(similar["c"])

                # More examples of the properties of detections:

                if detection_type == "visual.context" or detection_type == "audio.context":
                        if "ext_refs" in detection:
                                if "wikidata" in detection["ext_refs"]:
                                        print "Concept ID in Wikidata ontology: " + detection["ext_refs"]["wikidata"]["id"]
                                if "gkg" in detection["ext_refs"]:
                                        print "Concept ID in GKG ontology: " + detection["ext_refs"]["gkg"]["id"]

                if "occs" in detection:
                        for occ in detection["occs"]:
                                print "Occurrence starts at " + str(occ["ss"]) + "s from beginning of video, and ends at " + str(occ["se"]) + "s"
                                if "c_max" in occ:
                                        print "Maximum confidence of detection during this occurrence is " + str(occ["c_max"])
                                        # If you need the condifence for a particular time at second-level accuracy, see the by_second grouping of detections

                print
        print

# Example of listing only audio (speech) based word/phrase detections:
for detection_type, detections_of_this_type in metadata["detection_groupings"]["by_detection_type"].iteritems():
        if detection_type.startswith("audio.keyword."):
                for det_id in detections_of_this_type:
                        detection = metadata["detections"][det_id]
                        print "Label: " + detection["label"]  # etc... You get the idea :)
print

# Example of listing only detections of a specific detection type:
if "human.face" in metadata["detection_groupings"]["by_detection_type"]:
        for det_id in metadata["detection_groupings"]["by_detection_type"]["human.face"]:
                detection = metadata["detections"][det_id]  # etc...
print

# Example of listing IAB categories detected from different modalities (visual/audio/transcript) of the video
for detection_type, detections_of_this_type in metadata["detection_groupings"]["by_detection_type"].iteritems():
        if detection_type.startswith("topic.iab"):
                for det_id in detections_of_this_type:
                        detection = metadata["detections"][det_id]  # etc...
                        print "IAB label, simple: " + detection["label"]
                        print "IAB ID: " + detection["ext_refs"]["iab"]["id"]
                        print "IAB hierarchical label structure:"
                        print detection["ext_refs"]["iab"]["labels_hierarchy"]
print

# Time-based access: Loop over time (each second of the video) and access detections of each second
sec_index = -1
for secdata in metadata["detection_groupings"]["by_second"]:
        sec_index += 1
        print "----------"
        print "Detected at second " + str(sec_index) + ":"
        print
        for detdata in secdata:
                det_id = detdata["d"]
                if "c" in detdata:
                        print "At this second, detection has confidence " + str(detdata["c"])
                if "o" in detdata:
                        # If for some reason you need to know the corresponding occurrence (time-period that contains this second-based detection)
                        print "The detection at this second is part of one of more occurrences. The occurrence IDs, suitable for searching within the 'occs' list of the 'detection' object, are:"
                        for occ_id in detdata["o"]:
                                print occ_id
                print "Detection ID: " + det_id
                detection = metadata["detections"][det_id]
                print "Label: " + detection["label"]
                print "Detection of the type " + detection["t"] + ", full info:"
                # Of course, also here you can access attributes, cid, occurrences etc. through the "detection" object
                # just like when you listed detections by their type. In other words, when you just know the ID
                # of the detection, it's easy to read the information about the detection by using the ID.
                print detection
                print

How-to quick recipes

Analyze your video file (Quick Start)

Contact Valossa sales to activate a subscription relevant to your use case.
Get your confidential API key from your Valossa Portal profile page.
Analyze your video using the Valossa Core API or via the GUI in Valossa Portal.

Option 1: Using the Valossa Core API — Command line example

Create a JSON file containing the analysis request (see details for the new_job call in the API). A simple new_job example with only the necessary parameters:

{
  "api_key": "kkkkkkkkkkkkkkkkkkkk",
  "media": {
    "video": {
      "url": "https://example.com/my-vid.mpg"
    }
  }
}

Run command:

curl --header "Content-Type:application/json" -X POST -d @your_request.json https://api.valossa.com/core/1.0/new_job

Save the job_id which can be found in the 200 OK return response.
Your caller script may monitor the status of the job by using the job_status call in the API.
Using a callback URL is also possible, while not recommended. Instructions in the documentation of new_job (use search: Ctrl+F).

Option 2: Manual analysis via Valossa Portal GUI

Go to the Analyze page in Valossa Portal.
Upload the video file or use a URL of the video file.
You can view the progress on the Valossa Portal homepage.

Examine the results of the analysis

Option 1: Valossa Report

Browse to your Valossa Portal homepage to view the visualized report.

Option 2: Valossa Metadata Reader

Download the result metadata JSON using the job_results call in the API or from the Valossa Portal homepage.
Use Valossa Metadata Reader to quickly extract relevant detections from the metadata JSON file.

Option 3: Make your own parser

Download the result metadata JSON using the job_results call in the API or from the Valossa Portal homepage.
See the code examples for metadata reading. Get familiar with the basics of interpretation of Valossa Core metadata. The detections are there in the "detections" structure. A time-based, second-by-second index to the detections is given in the "by_second" structure, which references the actual detections by their detection ID strings. To list all detections of a specific type, you may use the "by_detection_type" structure.

Find video topics for advertising and content profiling

From the analysis results, pick the IAB advertising topics (or, if more suitable for you, the general topics or detailed audiovisual detections) so you can match the correct ads to your specific analyzed videos:

Contact Valossa sales for a subscription of Standard Metadata
Analyze your video
Get familiar with the basics of interpretation of Valossa Core metadata. The detections are there in the "detections" structure. A time-based, second-by-second index to the detections is given in the "by_second" structure, which references the actual detections by their detection ID strings. To list all detections of a specific type, you may use the "by_detection_type" structure.
Use the "topic.iab" and "topic.iab.section" detection types for IAB categories (or "topic.general" for topics that are non-IAB-based). See details

Extract visual-, audio-, and color-related tags

Pick relevant detections from the Valossa Core standard metadata:

Contact Valossa sales for a subscription of Standard Metadata
Analyze your video
Get familiar with the basics of interpretation of Valossa Core metadata. The detections are there in the "detections" structure. A time-based, second-by-second index to the detections is given in the "by_second" structure, which references the actual detections by their detection ID strings. To list all detections of a specific type, you may use the "by_detection_type" structure.
Read the results for the "visual.context" or "audio.context" or "visual.color" detection types. Please note: audio.context is for detecting various sounds, but in order to access the speech content (words, phrases) you need the speech-to-text related detections.

Extract all faces and recognize notable people

From the face-results, pick any matches of the Valossa Celebrities Gallery, or use nameless faces (no gallery matches) if suitable for your application, or create and use your custom face gallery:

Contact Valossa sales for a subscription of Standard Metadata or Face Analysis with Expressions and Voice Sentiment
Create custom face gallery/galleries if needed. You may skip this step if the Valossa Celebrities Gallery with its thousands of face identities is in itself enough for your application, or if your application doesn’t need face identities (i.e., needs only faces without names).
Analyze your video
Get familiar with the basics of interpretation of Valossa Core metadata. The detections are there in the "detections" structure. A time-based, second-by-second index to the detections is given in the "by_second" structure, which references the actual detections by their detection ID strings. To list all detections of a specific type, you may use the "by_detection_type" structure.
Read the results for the "human.face" detection type. Any gallery matches are provided in the "similar_to" fields. See details

Extract OCR results (recognizing visual text and numbers from videos)

OCR stands for Optical Character Recognition. Pick relevant detections from the Valossa Core standard metadata:

Contact Valossa sales for a subscription of Standard Metadata
Analyze your video
Get familiar with the basics of interpretation of Valossa Core metadata. The detections are there in the "detections" structure. A time-based, second-by-second index to the detections is given in the "by_second" structure, which references the actual detections by their detection ID strings. To list all detections of a specific type, you may use the "by_detection_type" structure.
Read the results for "visual.text_region.*" detection types. See details

Extract speech-to-text transcripts and generate subtitles

From the results, pick the speech-related detections, or download the speech-to-text content in the SRT or VTT formats:

Contact Valossa sales for a subscription of Standard Metadata or Face Analysis with Expressions and Voice Sentiment or Visual Video Moderation with the audio+speech addon
Analyze your video
Get familiar with the basics of interpretation of Valossa Core metadata. The detections are there in the "detections" structure. A time-based, second-by-second index to the detections is given in the "by_second" structure, which references the actual detections by their detection ID strings. To list all detections of a specific type, you may use the "by_detection_type" structure.
Read the results for the "audio.speech" detection type (see details), or "audio.speech_detailed" if your use case involves word-level time codes and your subscription has that specific technology enabled. (Additionally, the speech-to-text results converted into the SRT or VTT file formats are available for download in Valossa Portal, and via a job_results call with a special parameter.) If you need to access speech-based keywords, phrases and named entities, see below:

Extract keywords, phrases and named entities from speech

From the results, pick the audio-keywords-related detections:

Contact Valossa sales for a subscription of Standard Metadata or Face Analysis with Expressions and Voice Sentiment or Visual Video Moderation with the audio+speech addon
Analyze your video
Get familiar with the basics of interpretation of Valossa Core metadata. The detections are there in the "detections" structure. A time-based, second-by-second index to the detections is given in the "by_second" structure, which references the actual detections by their detection ID strings. To list all detections of a specific type, you may use the "by_detection_type" structure.
Read the results for the "audio.keyword.*" detection types (or, in case of your pre-existing transcripts having been analyzed, the results of the "transcript.keyword.*" detection types). See details

Analyze speech sentiment

This means sentiment from the content of the speech, that is, from the spoken words. From the results, pick the audio emotion related data:

Contact Valossa sales for a subscription of Face Analysis with Expressions and Voice Sentiment or Standard Metadata with the Facial expressions and Voice sentiment addon
Analyze your video
Get familiar with the basics of interpretation of Valossa Core metadata. The detections are there in the "detections" structure. A time-based, second-by-second index to the detections is given in the "by_second" structure, which references the actual detections by their detection ID strings. To list all detections of a specific type, you may use the "by_detection_type" structure.
Currently only valence from speech is supported for speech sentiment analysis. The valence information is available in the "audio.speech" detections.

Analyze content for sensitive or inappropriate content

Automatically filter your videos for inappropriate (sexual, violent, disturbing etc.) content from both visual, speech and audio detections. From the analysis results, check for the visual inappropriate detections in the content compliance category, inappropriate words or phrases from speech-to-text analysis, or inappropriate audio events from audio analysis:

Contact Valossa sales for a subscription of Visual Video Moderation with or without the audio+speech addon as applicable
Analyze your video
Get familiar with the basics of interpretation of Valossa Core metadata. The detections are there in the "detections" structure. A time-based, second-by-second index to the detections is given in the "by_second" structure, which references the actual detections by their detection ID strings. To list all detections of a specific type, you may use the "by_detection_type" structure.
If you are interested in visual matches, read the results for "visual.context" detections that have the "content_compliance" category tag (in addition, there are also other, more specific tags, such as "violence" or "sexual"). If you are interested in speech-based matches, read the results of the "audio.keyword.compliance" detections. If you are interested in sounds-based matches (such as explosions and grunts), read the results of the "audio.context" detections that have the "content_compliance" category tag.

Analyze emotions from faces

From the results, pick the face-sentiment related data in the time-based faces data:

Contact Valossa sales for a subscription of Face Analysis with Expressions and Voice Sentiment or Standard Metadata with the Facial expressions and Voice sentiment addon
Analyze your video
Get familiar with the basics of interpretation of Valossa Core metadata. The detections are there in the "detections" structure. A time-based, second-by-second index to the detections is given in the "by_second" structure, which references the actual detections by their detection ID strings. To list all detections of a specific type, you may use the "by_detection_type" structure.
Read the seconds-based results (in the "by_second" structure) for the "human.face" detections. The emotion information, if detected, is provided for each face identity in each 1-second slot. See details

Analyze audible emotions from voice

This means emotion from the audio prosodics such as laughter, NOT from the content (words) of speech itself. From the results, pick the voice emotion related data:

Contact Valossa sales for a subscription of Face Analysis with Expressions and Voice Sentiment or Standard Metadata with the Facial expressions and Voice sentiment addon
Analyze your video
Get familiar with the basics of interpretation of Valossa Core metadata. The detections are there in the "detections" structure. A time-based, second-by-second index to the detections is given in the "by_second" structure, which references the actual detections by their detection ID strings. To list all detections of a specific type, you may use the "by_detection_type" structure.
Read the results for the "audio.voice_emotion" detection type. See details

Automatically create trailer video clips (Autopreview)

The famous Valossa Autopreview that creates shorter preview video clips ("teaser videos", "summarization clips") from your videos. Those clips are valuable in, for example, a video-on-demand service. To learn more about Valossa Autopreview, first contact sales.

Contact Valossa sales for a subscription of Autopreview for Automated Video Highlights Production
Analyze your video
Get familiar with the basics of interpretation of Valossa Core metadata. The detections are there in the "detections" structure. A time-based, second-by-second index to the detections is given in the "by_second" structure, which references the actual detections by their detection ID strings. To list all detections of a specific type, you may use the "by_detection_type" structure.
Engage in a session with the friendly Valossa staff to learn more about how to configure your Autopreview to suit better your needs. When the correct configuration is active, each video analysis will produce one or more correctly parametrized Autopreview video clip(s). Downloading the clips automatically into your production system involves typically an integration step, where the friendly Valossa staff will collaborate with your technical contact person.

Valossa Core metadata JSON format version changelog

1.5.2: added new Valossa category tags for topic.iab and topic.iab.section, updated the IAB taxonomy version number to 2.2 in those detection types

1.5.1: added SCD support to IAB detections that have the IAB taxonomy version 2.1

1.5.0: voice emotion as part of seconds-based audio.speech data deprecated and moved into the new dedicated detection type audio.voice_emotion

1.4.3: topic.iab.section added, voice emotion added to seconds-based data of audio.speech, human.face supports additional information about an identity

1.4.2: audio.speech_detailed added

1.4.1: visual.text_region.full_frame_analysis, visual.text_region.lower_third, visual.text_region.middle_third, visual.text_region.upper_third added

1.4.0: some tag categories had their identifiers changed

1.3.12: face height per second added

1.3.11: by_frequency, topic.iab, topic.general added

1.3.10: shs added to occurrences

1.3.9: n_audio_channels added to technical media information

1.3.8: visual.object.localized added

1.3.7: changed representation of bitrate (bps) in technical media information

1.3.6: resolution, codecs and bitrates added to technical media information

1.3.5: visual.color added, violence-related concept categories added

1.3.4: detection grouping by_detection_property added, identifier information added for gallery faces

1.3.3: categ added to relevant visual.context and audio.context detections

1.3.2: similar_to in human.face detections supports role names

1.3.1: added metadata type field (supports distinguishing between different types of Valossa metadata in the future)

1.3.0: improved speech-to-text format

1.2.1: speech-to-text

1.2.0: field naming improved

1.1.0: more compact format

1.0.0: large changes, completely deprecated old version 0.6.1.

Thousands of recognized people (faces) in the Valossa Celebrities Gallery

This face gallery contains thousands of globally notable celebrities.

The number of faces available has been multiplied in 2021, however not sacrificing the quality of face choices. The list has been carefully curated and combined with practical information related to the personal identities.

If you have questions about the face gallery or about integrating our efficient and easy-to-use face recognition solution into your application, contact sales to initiate discussion.

You can teach Valossa AI to recognize more people by using the Training API, which allows you to create your custom galleries that are used in your video analyses (alongside the Valossa Celebrities Gallery, which is always included in a face analysis).

Recognized logos

The logos available are a selection of top global brands.

Questions? Wishes for a custom set of logos, or maybe desires to detect some other visual objects with their bounding box coordinates provided like the logos already have? Contact sales.