Reproduce “Online Generic Event Boundary Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hyungrok Jung, Daneul Kim, Seunggyun Lim, Jeany Son, Jonghyun Choi
Researchers from GIST, Seoul National University, and POSTECH introduce Online Generic Event Boundary Detection (On-GEBD) and the ESTimator model, which uses principles from human cognition to identify taxonomy-free event boundaries in real-time video streams. ESTimator outperforms existing online baselines, achieving an average F1 score of 0.748 on Kinetics-GEBD, and demonstrates strong zero-shot generalization to unseen datasets.
Key insights
- Leveraging cognitive science principles, specifically Event Segmentation Theory, to design AI models leads to more robust and human-aligned online video understanding systems.
- A dynamic thresholding mechanism for boundary detection, which adapts to the immediate temporal context by normalizing current prediction error against recent error history, is crucial for effective online performance.
- The inclusion of 'outlier' errors (from previously detected boundaries) in the memory queue significantly enhances the adaptive capability of the Online Boundary Discriminator, improving its sensitivity to new, meaningful changes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.