稀疏时间池网络的弱监督动作定位.pdf

Weakly Supervised Action Localization by Sparse Temporal Pooling Network 8 1 Abstract 0 2 RGB Action Stream Class r We propose a weakly supervised temporal action local- p ization algorithm on untrimmed videos using convolutional A neural networks. Our algorithm learns from video-level RGB T-CAM class labels and predicts temporal intervals of human ac- 3 tions with no requirement of temporal localization annota- Video Proposals Action ] tions. We design our network to identify a sparse subset of Location V key segments associated with target actions in a video us- FLOW T-CAM C ing an attention module and fuse the key segments through s. adaptive temporal pooling. Our loss function is comprised c of two terms that minimize the video-level action classifica- FLOW [ tion error and enforce the sparsity of the segment selection. Stream 2 At inference time, we extract and score temporal proposals v using temporal class activations and class-agnostic atten- 0 tions to estimate the time intervals that correspond to tar- Figure 1: Overview of the proposed algorithm. Our algorithm 8 get actions. The proposed algorithm attains state-of-the-art takes a two-stream input—RGB frames and optical flow between 0 results on the THUMOS14 dataset and outstanding perfor-