LkSFocalNets: Video Action Recognition With Large Kernel Selective Focal Networks

Abstract

Video action recognition tasks face the trade-off challenge between computational cost and performance. Existing models often compromise their ability to capture extensive contextual information to reduce computational complexity. To address this issue, this paper proposes Large Kernel Selective Focal Networks (LkSFocalNet), an efficient network architecture. First, we propose the Kernel Selective Block (KSB) and the Video Focal Block (VFB). The KSB dynamically adjusts the receptive field to capture essential contextual information, while the VFB models spatio-temporal dependencies through video focus modulation. Additionally, we present the Depthwise Separable Convolution Gated Unit (DsCG), which reduces computational load while enhancing the network’s robustness. Finally, these three blocks are combined into a Large Kernel Selective Focal Module (LkSFM), which is stacked to form LkSFocalNet. Experimental results on four benchmark datasets demonstrate that LkSFocalNet outperforms baseline models and achieves state-of-the-art performance on the Diving-48 dataset, all while maintaining lower computational complexity.