Temporal localization of mature content in long-form videos using only video-level labels

1. A computer-implemented method comprising:

receiving a request to train a machine learning model on a training video file comprising at least one mature content label;

training the machine learning model to generate a feature vector for each of a plurality of video frames of the training video file, generate a plurality of frame-level mature content classification scores of the training video file from the feature vectors of the training video file, and generate a video-level mature content classification score of the training video file from the plurality of frame-level mature content classification scores for the training video file based at least in part on the at least one mature content label of the training video file;

receiving a request for an input video file;

generating, by the machine learning model in response to the request:

a feature vector for each of a plurality of video frames of the input video file,

a plurality of frame-level mature content classification scores of the input video file from the feature vectors of the input video file by sending the feature vector to a first branch of the machine learning model that determines a class-agnostic attention value for each of the plurality of video frames of the input video file, and a second branch of the machine learning model that determines a binary classification score for each of the plurality of video frames of the input video file, and

a video-level mature content classification score of the input video file from the plurality of frame-level mature content classification scores for the input video file that comprises a sum of each resultant of a multiplication of the class-agnostic attention value for a frame of the plurality of video frames of the input video file by the binary classification score of the frame; and

transmitting the plurality of frame-level mature content classification scores of the input video file or the video-level mature content classification score of the input video file to a client application or to a storage location.

2. The computer-implemented method of claim 1 , wherein the training the machine learning model to generate the feature vector for each of a plurality of video frames of the training video file comprises generating the feature vector that comprises video features and audio features for corresponding audio of the plurality of video frames of the training video file.

3. The computer-implemented method of claim 1 , further comprising:

displaying, via a graphical user interface to a prospective viewer of the input video file, a skip query for a scene of the input video file having at least one frame-level mature content classification score that exceeds a score threshold;

receiving an indication from the prospective viewer via the graphical user interface to skip the scene; and

causing skipping of the scene to a following scene when the input video file is viewed by the prospective viewer.

4. A computer-implemented method comprising:

receiving a content detection request for a video file;

generating, by a machine learning model in response to the content detection request:

a feature vector for each of a plurality of video frames of the video file,

a plurality of frame-level content classification scores of the video file from the feature vectors of the video file by sending the feature vector to a first branch of the machine learning model that determines a class-agnostic attention value for each of the plurality of video frames of the video file, and a second branch of the machine learning model that determines a binary classification score for each of the plurality of video frames of the video file, and

a video-level content classification score of the video file from the plurality of frame-level content classification scores for the video file that comprises a sum of each resultant of a multiplication of the class-agnostic attention value for a frame of the plurality of video frames of the video file by the binary classification score of the frame; and

transmitting the plurality of frame-level content classification scores of the video file or the video-level content classification score of the video file to a client application or to a storage location.

5. The computer-implemented method of claim 4 , wherein the generating the feature vector for each of the plurality of video frames of the video file comprises generating the feature vector that comprises video features and audio features for corresponding audio of the plurality of video frames of the video file.

6. The computer-implemented method of claim 5 , where the video features are concatenated with the audio features in the feature vector.

7. The computer-implemented method of claim 4 , wherein the classification scores indicate a likelihood of sexually explicit content.

8. The computer-implemented method of claim 4 , further comprising:

displaying, via a graphical user interface to a viewer of the video file, a skip query for a scene of the video file having at least one frame-level content classification score that exceeds a score threshold;

receiving an indication from the viewer via the graphical user interface to skip the scene; and

causing skipping of the scene to a following scene when the video file is viewed by the viewer.

9. The computer-implemented method of claim 8 , wherein the causing the skipping of the scene comprises removing the scene from the video file before delivering at least a corresponding subset of the video file to the viewer.

10. The computer-implemented method of claim 8 , wherein the displaying the skip query begins after viewing of the video file by the viewer has begun.

11. The computer-implemented method of claim 4 , further comprising

displaying, via a graphical user interface, an indication of the video file when the video-level content classification score exceeds a score threshold.

12. The computer-implemented method of claim 4 , further comprising displaying, via a graphical user interface, an indication of a scene of the video file having the at least one frame-level content classification score that exceeds a score threshold.

13. A system comprising:

one or more electronic devices comprising a processor to implement a storage service to store a video file; and

one or more electronic devices comprising a processor to implement a content detector service, the content detector service including instructions that upon execution cause the content detector service to perform a method comprising:

receiving a content detection request for the video file;

generating, by a machine learning model in response to the content detection request:

a feature vector for each of a plurality of video frames of the video file,

a plurality of frame-level content classification scores of the video file from the feature vectors of the video file by sending the feature vector to a first branch of the machine learning model that determines a class-agnostic attention value for each of the plurality of video frames of the video file, and a second branch of the machine learning model that determines a binary classification score for each of the plurality of video frames of the video file, and

a video-level content classification score of the video file from the plurality of frame-level content classification scores for the video file that comprises a sum of each resultant of a multiplication of the class-agnostic attention value for a frame of the plurality of video frames of the video file by the binary classification score of the frame; and

transmitting the plurality of frame-level content classification scores of the video file or the video-level content classification score of the video file to a client application or to a storage location.

14. The system of claim 13 , wherein the instructions upon execution cause the content detector service to perform operations wherein the generating the feature vector for each of the plurality of video frames of the video file comprises generating the feature vector that comprises video features and audio features for corresponding audio of the plurality of video frames of the video file.

15. The system of claim 14 , wherein the video features are concatenated with the audio features in the feature vector.

16. The system of claim 13 , wherein the instructions upon execution cause the content detector service to further perform operations comprising:

displaying, via a graphical user interface to a viewer of the video file, a skip query for a scene of the video file having at least one frame-level classification score that exceeds a score threshold;

receiving an indication from the viewer via the graphical user interface to skip the scene; and

causing skipping of the scene to a following scene when the video file is viewed by the viewer.

17. The system of claim 16 , wherein the displaying the skip query begins after viewing of the video file by the viewer has begun.

18. The system of claim 13 , wherein the instructions upon execution cause the content detector service to further perform operations comprising displaying, via a graphical user interface, an indication of the video file when the video-level content classification score exceeds a score threshold.

19. The system of claim 13 , wherein the instructions upon execution cause the content detector service to further perform operations comprising displaying, via a graphical user interface, an indication of a scene of the video file having at least one frame-level content classification score that exceeds a score threshold.

20. The system of claim 13 , wherein the classification scores indicate a likelihood of sexually explicit content.