CodecMotion-OAD: Efficient Online Action Detection with Compressed-Domain Motion
Not provided in the abstract
Abstract
This paper presents CodecMotion-OAD, an efficient online action detection method that utilizes codec motion vectors to reduce computational bottlenecks associated with optical flow extraction in streaming videos.
Reality Card
CodecMotion-OAD achieves competitive action detection performance while significantly reducing processing time compared to traditional optical flow methods.
Achieves 68.3 mAP on THUMOS'14 while reducing decoder-to-prediction processing time by 11.48-fold on CPU and 2.04-fold on GPU.
The reliance on pretrained flow-feature interfaces without action labels may limit generalizability to other datasets.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.