Improving Viewpoint-Invariance and Temporal Consistency for Action Detection
Researchers have developed a new two-stage approach to improve human action detection in untrimmed videos. The method enhances viewpoint invariance by extracting motion features from augmented virtual viewpoints during training. A subsequent stage employs a novel view-invariant temporal encoder, utilizing selective state-space sequence modeling, to integrate information across different viewpoints and time scales. This technique has demonstrated superior performance on the PKU-MMD and BABEL benchmarks compared to existing state-of-the-art methods. AI
IMPACT Enhances action detection capabilities, potentially improving applications in video analysis and surveillance.