Understanding what individuals are doing in crowded, multi-person video is important for real-world systems, but current benchmarks may not capture that task. A new arXiv preprint argues that existing datasets often simplify the problem into grounded group activity recognition, which loses the individual-level detail.
The paper frames individual characteristics and activities as the core challenge, rather than treating a whole group as a single unit. That distinction matters for applications that need to track or describe specific people in complex scenes.
The source abstract is truncated, so the article does not describe the proposed method, experiments, or results. Readers looking for those details will need the full paper.