Online action detection (OAD) aims to identify actions as they happen in streaming video. Many current systems combine RGB appearance with optical flow to capture motion, but extracting dense optical flow is a major computational bottleneck that limits real-time performance.
A new arXiv preprint, CodecMotion-OAD, proposes an alternative: use motion information already embedded in compressed video streams, such as codec motion vectors. This compressed-domain motion can substitute for dense optical flow, potentially making OAD much more efficient.
The abstract does not provide quantitative results, but the motivation is clear: by avoiding dense-flow extraction, the method targets a key inefficiency in existing OAD pipelines. The work is described as an efficient approach, though details on architecture and benchmarks remain limited in the announcement.