A Tracking-Based Two-Stage Framework for Spatio-Temporal Action Detection

Spatio-temporal action detection (STAD) is a task receiving widespread attention and has numerous application scenarios, such as video surveillance and smart education. Current studies follow a localization-based two-stage detection paradigm, which exploits a person detector for action localization and a feature processing model with a classifier for action classification. However, many issues occur due to the imbalance between task settings and model complexity in STAD. Firstly, the model complexity of heavy offline person detectors adds to the inference overhead. Secondly, the frame-level actor proposals are incompatible with the video-level feature aggregation and Region-of-Interest feature pooling in action classification, which limits the detection performance under diverse action motions and results in low detection accuracy. In this paper, we propose a tracking-based two-stage spatio-temporal action detection framework called TrAD. The key idea of TrAD is to build video-level consistency and reduce model complexity in our STAD framework by generating action track proposals among multiple video frames instead of actor proposals in a single frame. In particular, we utilize tailored tracking to simulate the behavior of human cognitive actions and used the captured motion trajectories as video-level proposals. We then integrate a proposal scaling method and a feature aggregation module into action classification to enhance feature pooling for detected tracks. Evaluations in the AVA dataset demonstrate that TrAD achieves SOTA performance with 29.7 mAP, while also facilitating a 58% reduction in overall computation compared to SlowFast.