Detect Video End Credits

This a hack-day experiment to programmatically detect the approximate timestamp of the end credits for a given video.

The end credit is a particularly useful timestamp for video players to show additional content in the UI. For example, Netflix will displays a piece of UI to navigate to the next episode when the end credits starts playing for the current video.

Running

To run the project, curl, docker and docker-compose are required.

./scripts/setup.sh
docker-compose build
docker-compose run main

Dependencies

This demo takes advantage of the following core dependencies.

• OpenCV - video analysis
• Tesseract - optical character recognition

Approach

The rough approach is to:

1. Capture frames for the video for some interval e.g. 1 frame per second.
2. Score each frame for how likely it is a end credit frame.
3. Analyze the pattern of scores to determine the beginning of the end credit.

Heuristics

To score the likelihood that a given frame is a end credit frame, we can use the following heuristics.

1. Whether certain keywords are present. e.g. "Produced By", "Director", "Dolby".
2. How many lines of text are present or text density end credit screen likely contain dense text count.
3. The background color of the frame since end credit screen are often (though not always) black.