Reproduce “Generic Event Boundary Detection: A Benchmark for Event Segmentation” 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
Mike Zheng Shou, Stan Weixian Lei, Weiyao Wang, Deepti Ghadiyaram, Matt Feiszli
Researchers from Facebook AI and NUS define Generic Event Boundary Detection (GEBD) as a taxonomy-free task for segmenting long-form videos into human-perceived events, introducing the large-scale Kinetics-GEBD benchmark. Benchmarking reveals that a new supervised method (PC) and a cognitively-inspired unsupervised method (PA) perform strongly, with GEBD segments improving video summarization and enabling efficient classification.
Key insights
- Despite the taxonomy-free nature of the task, human annotators showed a 'decent degree of consensus' on event boundaries, validating the annotation guidelines and the naturalness of the GEBD concept.
- Cognitively-inspired unsupervised methods, such as PredictAbility (PA), achieved competitive performance on event boundary detection, sometimes surpassing previous supervised methods on the TAPOS dataset, by identifying moments of least predictability.
- Generic event boundaries provide meaningful temporal units that are highly valuable for downstream tasks, improving video summarization and offering an efficient, data-driven strategy for video-level classification.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.