Action Temporal Localization in Untrimmed Videos via Multi-stage CNNs

We address action temporal localization in untrimmed long videos. This is important because videos in real applications are usually unconstrained and contain multiple action instances plus video content of background scenes or other activities. To address this challenging issue, we exploit the effectiveness of deep networks in action temporal localization via multi-stage segment-based 3D ConvNets: (1) a proposal stage identifies candidate segments in a long video that may contain actions; (2) a classification stage learns one-vs-all action classification model to serve as initialization for the localization stage; and (3) a localization stage fine-tunes on the model learnt in the classification stage to localize each action instance. We propose a novel loss function for the localization stage to explicitly consider temporal overlap and therefore achieve high temporal localization accuracy. On two large-scale benchmarks, our approach achieves significantly superior performances compared with other state-of-the-art systems: mAP increases from 1.7% to 7.4% on MEXaction2 and increased from 15.0% to 19.0% on THUMOS 2014, when the overlap threshold for evaluation is set to 0.5.