Implement computer vision solutions 3.2 Design and implement multimodal understanding workflows
Video editors use a custom Content Understanding video analyzer and need the exact…
Study guide skill: Implement video analysis workflows to process and interpret video segments
Question B5-006
Video editors use a custom Content Understanding video analyzer and need the exact timestamps where camera cuts occur, so they can repackage clips at existing edit points. The output shows key frames but no cut timestamps. What should you change?
Select one answer.
Answer: A
Shot detection is part of content extraction: it returns the start times of camera shots as cameraShotTimesMs, and that output is included only when returnDetails is true. Key frames are sampled from within each shot, so they aren't cut points. Custom segmentation has the model create segments that match a natural-language description, while shot detection is deterministic and aligned with existing edits; turning returnDetails off would remove the camera shot output.
Deep dive
How it works
Content Understanding video analysis starts with a content extraction pass that finds who's speaking and where the cuts are. Cuts come from shot detection: it identifies segments aligned with shot boundaries where possible, so you can edit and repackage content with breaks exactly at existing edits. The output is a list of millisecond timestamps in cameraShotTimesMs.
The catch is that this output appears only when returnDetails is true in the analyzer definition. The video elements table confirms it: camera shots require returnDetails, while key frames don't. That's why the editors see key frames but no cut timestamps. Set returnDetails to true and re-run.
Each cameraShotTimesMs value marks where camera shots change, indicating cuts, transitions or significant changes in camera angle, as milliseconds from the start of the video.
The other options, and when you'd use them
- enableSegment with cuts described in contentCategories: That's custom segmentation, where a generative model splits the video into logical chunks based on categories you define. It's semantic, not aligned to the existing edit points.
- Use keyFrameTimesMs as the cut points: Key frames are sampled uniformly from within each camera shot to feed field extraction. They sit inside shots, not on the boundaries.
- Set returnDetails to false because shots are always returned: Backwards. Shot detection output is only returned when returnDetails is true.
Exam tip
Exact cut timestamps from Content Understanding video = cameraShotTimesMs, which needs returnDetails: true. Key frames are inside shots, not at cuts.
References
Last reviewed: