Create AI CM:hard_nudity task

This algorithm allows to detect explicit nudity of the human body (involving genitals) in a video. Generic info about all capabilities and limits see in the generic “Content Moderation” method. What is Hard nudity detection? This method is often used to analyze UGC to determine whether videos can be published to all users, or to prohibit publication due to offensive and inappropriate content. Objects that can be detected:

• ANUS_EXPOSED
• BUTTOCKS_EXPOSED
• FEMALE_BREAST_EXPOSED
• FEMALE_GENITALIA_EXPOSED
• MALE_BREAST_EXPOSED
• MALE_GENITALIA_EXPOSED Please note that the number of objects is less than in the soft-nudity-detection. This method works faster and better if only exposed body parts detection is required. How to use? The information is returned with the video frame number where it was found and probability of the detected object. Nudity detection is done using AI, so for each object a probability percentage is applied; objects with a probability of at least 30% are included in the response. Video processing speed is approximately 1:5. Example of detected nudity or body parts:

{
  "nudity_detected": true,
  "detection_results": [ "MALE_GENITALIA_EXPOSED" ]
  "frames": [
      {
          "confidence": 0.75,
          "frame_number": 35,
          "label": "MALE_GENITALIA_EXPOSED"
      },...
  ]
}

Example response when nudity or body parts were not found:

{
  "nudity_detected": false,
  "detection_results": []
  "frames": []
}

There is no universal recipe under which a video can be considered unacceptable, since different services host different types of videos for different audiences: adult content, children’s content, educational content, etc. You can determine the probability threshold at which you consider a video inappropriate. The easiest option is to run several of your videos and analyze the resulting probability coefficient.

Sometimes a detected object at the beginning of the video immediately makes it clear that there is no need to further analyze the video. For such cases, you can use stop tags. Use parameter “stop_objects” to specify comma separated stop tags. It is also possible to specify % probability threshold value, above which the stop tag will be triggered.

  {
      "url": "...",
      "stop_objects": "MALE_GENITALIA_EXPOSED:0.8,FEMALE_GENITALIA_EXPOSED"
  }

Please note that the API only provides a set of data (json) about the objects found, so no video is generated. The demo video video (above ^) was specially created based on json from the API for visual demonstration and better perception of the possibilities.