
科学技术与工程

中国技术经济学会
Dual-path Video Action Recognition Method Fusing Dynamic Token Sparsification and Cross-path Attention
Jun-xin LUO , Long-ye WANG , Xiao-li ZENG , Xing-yu DENG , Xin LAN
Science Technology and Engineering ›› 2026, Vol. 26 ›› Issue (31) : 13546 -13554.
To enhance the computational efficiency and cross-path feature interaction of the SlowFast network in action recognition tasks, a video dynamic sparse token (VDST) selection module and a bidirectional gated cross-attention module (Bi-CAM) are incorporated into the original dual-path framework. By adaptively selecting key spatiotemporal regions in the fast pathway and establishing fine-grained bidirectional semantic fusion between the slow and fast pathways, the trade-off between computational complexity and recognition performance in the improved SlowFast architecture is investigated. It is shown that redundant feature computation is effectively reduced and model accuracy is maintained by the VDST module while FLOPs(floating point operations) are significantly lowered; semantic interaction across pathways is enhanced and more comprehensive action representations are obtained by the Bi-CAM module; and a top-1 accuracy of 95.5% on the UCF101 dataset with only 33.6 GFLOPs(giga floating point operations) is achieved by the combined model. It is demonstrated that computational efficiency is substantially improved and high recognition accuracy is preserved by the proposed multi-module enhanced SlowFast framework, providing a feasible solution for efficient video understanding.
video action recognition / SlowFast network / dynamic sparse Tokens / gating mechanism / bidirectional cross-attention
/
| 〈 |
|
〉 |