Dilated Multi-Temporal Modeling for Action Recognition

Action recognition involves capturing temporal information from video clips where the duration varies with videos for the same action. Due to the diverse scale of temporal context, uniform size kernels utilized in convolutional neural networks (CNNs) limit the capability of multiple-scale temporal modeling. In this paper, we propose a novel dilated multi-temporal (DMT) module that provides a solution for modeling multi-temporal information in action recognition. By using dilated convolutions with different dilation rates in different feature map channels, the DMT module captures information at multiple scales without the need for costly multi-branch networks, input-level frame pyramids, or feature map stacking that previous works have usually incurred. Therefore, this approach enables the integration of temporal information from multiple scales. In addition, the DMT module can be integrated into existing 2D CNNs, making it a straightforward and intuitive solution for addressing the challenge of multi-temporal modeling. Our proposed method has demonstrated promising results in performance and has achieved about 2% and 1% accuracy improvement on FineGym99 and SthV1. We conducted an empirical analysis that demonstrates how DMT improves the classification accuracy for action classes with varying durations.