Breakthrough
New Detection Methods
Parent tracking issue for discussion and development of new detection methods.
This tracking issue can also be used for improvements. As an example this paper uses ContentDetector with a threshold of 20, followed by another post-processing stage:
Many videos remained unclipped by PySceneDetect, despite clear action or event transitions, due to its reliance on HSV-based frame differences that miss semantic changes. To overcome this, we extract SigLIP [69] frame embeddings at 3 FPS and apply a custom clipping algorithm using Hierarchical Agglomerative Clustering to segment semantically distinct video segments.
It's unclear if we can apply this same principle in the absence of dataset labels, but it's worth exploring. This is similar in theory to the approach taken by #441 in that it considers more semantic changes around the cut rather than just frame-by-frame differences. Determining how to best integrate approaches like that will be really beneficial in improving algorithm robustness going forwards.