Global-local Temporal Modeling Enhanced 2DCNN with Category-supervised Contrastive Learning for Action Recognition

Action recognition is a core fundamental task in computer vision, with broad applications in critical scenarios such as security monitoring and autonomous driving. However, previous 2DCNN action recognition approaches usually suffer from the following problems: insufficient temporal modeling capability as well as unclear demarcation of semantic boundaries for confusable categories. To tackle these issues, we devise a novel Global-local Temporal Modeling Enhanced 2DCNN with Category-supervised Contrastive Learning (GCNet), consisting of a Global-local Temporal Model module (GTM) and a Category-supervised Contrastive Learning (CCL). Specifically, the Global-local Temporal module constructs an overall semantic model of the action by analyzing the trend of feature changes over time. The Category-supervised Contrastive Learning module, on the other hand, makes the semantic boundaries of actions clearer by comparing positive and negative samples. Extensive experiments conducted on the widely used action recognition datasets (UCF-101, HMDB-51, Something Something-v1), have demonstrated that our proposed GCNet can achieve the best performance compared to other state-of-the-art approaches. We release the code at https://github.com/luoyujuan123/GCNet.