TAAD
Spatio-Temporal Action Detection Under Large Motion[1]
作者是来自ETHZ的Gurkirt Singh, Vasileios Choutas, Suman Saha, Fisher Yu和Luc Van Gool。论文引用[1]:Singh, Gurkirt et al. “Spatio-Temporal Action Detection Under Large Motion.” 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2022): 5998-6007.
Time
- 2022.Oct
Key Words
- track information for feature aggregation rather than tube from proposals
- 3 motion categories: large motion、medium motion、small motion
总结
- 当前的STAD的tube detection的方法经常将一个给定的keyframe上的bbox proposal扩展成一个3D temporal cuboid,然后从邻近帧进行pool features。如果actor的位置或者shape表现出了large 2D motion和variability through frames,这样的pooling不能够积累有意义的spaito-temporal features。在这个工作中,作者旨在研究cuboid-aware feature aggregation in action detection under large action。进一步,提出了在large motion的情况下,通过tracking actors和进行temporal feature aggregation along the respective tracks增强actor feature representation,定义了在不同的固定的time scales下的actor motion的IoU。有large motion的action会随着时间导致lower IoU,slower actions会随着时间维持higher IoU。作者发现track-aware feature aggregation持续地实现了很大的提升in action detection。