TadML: A Fast Temporal Action Detection with Mechanics-MLP

Temporal Action Detection (TAD) involves identifying action categories and their respective start and end frames in lengthy untrimmed videos, with current models utilizing both RGB and optical flow streams that require manual intervention, add computational complexity, and consume time. Moreover, two-stage approaches prioritizing proposal generation in the ini-tial stage result in a substantial reduction in inference speed. To address this, we propose a single-stage anchor-free method that solely utilizes the RGB stream and incorporates a novel Newtonian Mechanics-MLP architecture. Our model achieves comparable accuracy to existing state-of-the-art models but with significantly faster inference speeds, clocking in at an av-erage of 4.44 videos per second on THUMOS14. Our approach showcases the potential of MLP in downstream tasks like TAD. The source code is available at https://github.com/BonedDeng/TadML .