One major challenge of Weakly-supervised Temporal Action Localization (WTAL) is to handle diverse backgrounds in videos. To model background frames, most existing methods treat them as an additional action class. However, because background frames usually do not share common semantics, squeezing all the different background frames into a single class hinders network optimization. Moreover, the network would be confused and tends to fail when tested on videos with unseen background frames. To address this problem, we propose an Entropy Guided Attention Network (EGA-Net) to treat background frames as out-of-domain samples. Specifically, we design a two-branch module, where a domain branch detects whether a frame is an action by learning a class-agnostic attention map, and an action branch recognizes the action category of the frame by learning a class-specific attention map. By aggregating the two attention maps to model the joint domain-class distribution of frames, our EGA-Net can handle varying backgrounds. To train the class-agnostic attention map with only the video-level class labels, we propose an Entropy Guided Loss (EGL), which employs entropy as the supervision signal to distinguish action and background. Moreover, we propose a Global Similarity Loss (GSL) to enhance the action-specific attention map via action class center. Extensive experiments on THUMOS14, ActivityNet1.2 and ActivityNet1.3 datasets demonstrate the effectiveness of our EGA-Net.
Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)
This research / project is supported by the Agency for Science, Technology and Research - AME Programmatic Funding
Grant Reference no. : A18A2b0046