Temporal Inconsistency in Action Recognition for Long-Duration Video Streams

Abstract

Video-based action recognition systems are widely used in sports analysis, safety monitoring, and behavioral research. Most models are trained on short, curated clips that poorly reflect continuous real-world streams. This paper investigates temporal inconsistency in action predictions under long-duration recordings. Experiments show frequent boundary confusion when activities overlap or occur intermittently. Simple temporal smoothing improves visual continuity but reduces responsiveness. More principled sequence modeling remains necessary for stable deployment.