Early Action Recognition With Category Exclusion Using Policy-Based Reinforcement Learning

摘要: The goal of early action recognition is to predict label when the sequence partially observed. existing methods treat task as sequential classification problems on different observation ratios an sequence. Since these models are trained by differentiating positive category from all negative classes, diverse information categories ignored, which we believe can be collected help improve performance. In this paper, step towards a new direction introducing exclusion recognition. We model mask operation probability output pre-trained classifier. Specifically, use policy-based reinforcement learning train agent. agent generates series binary masks exclude interfering during execution and hence accuracy. proposed method evaluated three benchmark datasets, NTU-RGBD, First-Person Hand Action, well UCF-101. enhances accuracy consistently over where improvements stages especially significant.