A Multitemporal Scale and Spatial–Temporal Transformer Network for Temporal Action Localization

Temporal action localization plays an important role in video analysis, which aims to localize and classify actions in untrimmed videos. Previous methods often predict actions on a feature space of a single temporal scale. However, the temporal features of a low-level scale lack sufficient semantics for action classification, while a high-level scale cannot provide the rich details of the action boundaries. In addition, the long-range dependencies of video frames are often ignored. To address these issues, a novel multitemporal-scale spatial–temporal transformer (MSST) network is proposed for temporal action localization, which predicts actions on a feature space of multiple temporal scales. Specifically, we first use refined feature pyramids of different scales to pass semantics from high-level scales to low-level scales. Second, to establish the long temporal scale of the entire video, we use a spatial–temporal transformer encoder to capture the long-range dependencies of video frames. Then, the refined features with long-range dependencies are fed into a classifier for coarse action prediction. Finally, to further improve the prediction accuracy, we propose a frame-level self-attention module to refine the classification and boundaries of each action instance. Most importantly, these three modules are jointly explored in a unified framework, and MSST has an anchor-free and end-to-end architecture. Extensive experiments show that the proposed method can outperform state-of-the-art approaches on the THUMOS14 dataset and achieve comparable performance on the ActivityNet1.3 dataset. Compared with A2Net (TIP20, Avg{0.3:0.7}), Sub-Action (CSVT2022, Avg{0.1:0.5}), and AFSD (CVPR21, Avg{0.3:0.7}) on the THUMOS14 dataset, the proposed method can achieve improvements of 12.6%, 17.4%, and 2.2%, respectively.