Weakly-supervised Temporal Action Localization by Uncertainty Modeling

Weakly-supervised temporal action localization aims to learn detecting intervals of classes with only video-level labels. To this end, it is crucial separate frames from the background (i.e., not belonging any classes). In paper, we present a new perspective on where they are modeled as out-of-distribution samples regarding their inconsistency. Then, can be detected by estimating probability each frame being out-of-distribution, known uncertainty, but infeasible directly uncertainty without frame-level realize learning in weakly-supervised setting, leverage multiple instance formulation. Moreover, further introduce entropy loss better discriminate encouraging in-distribution (action) probabilities uniformly distributed over all classes. Experimental results show that our modeling effective at alleviating interference and brings large performance gain bells whistles. We demonstrate model significantly outperforms state-of-the-art methods benchmarks, THUMOS'14 ActivityNet (1.2 & 1.3). Our code available https://github.com/Pilhyeon/WTAL-Uncertainty-Modeling.