ACGNet: Action Complement Graph Network for Weakly-Supervised Temporal Action Localization

Weakly-supervised temporal action localization (WTAL) in untrimmed videos has emerged as a practical but challenging task since only video-level labels are available. Existing approaches typically leverage off-the-shelf segment-level features, which suffer from spatial incompleteness and incoherence, thus limiting their performance. In this paper, we tackle problem new perspective by enhancing representations with simple yet effective graph convolutional network, namely complement network (ACGNet). It facilitates the current video segment to perceive spatial-temporal dependencies others that potentially convey complementary clues, implicitly mitigating negative effects caused two issues above. By means, features more discriminative robust variations, contributing higher accuracies. More importantly, proposed ACGNet works universal module can be flexibly plugged into different WTAL frameworks, while maintaining end-to-end training fashion. Extensive experiments conducted on THUMOS'14 ActivityNet1.2 benchmarks, where state-of-the-art results clearly demonstrate superiority of approach.