Entropy guided attention network for weakly-supervised action localization

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 frames usually do not share common semantics, squeezing all the different into a single class hinders network optimization. Moreover, would be confused and tends fail when tested on videos with unseen frames. address this problem, we propose Entropy Guided Attention Network (EGA-Net) out-of-domain samples. Specifically, design two-branch module, where domain branch detects whether frame by learning class-agnostic attention map, recognizes category class-specific map. By aggregating two maps joint domain-class distribution our EGA-Net can varying backgrounds. train map only video-level labels, Loss (EGL), which employs entropy supervision signal distinguish background. Global Similarity (GSL) enhance action-specific via center. Extensive experiments THUMOS14, ActivityNet1.2 ActivityNet1.3 datasets demonstrate effectiveness EGA-Net.