Temporal Similarity-Based Computation Reduction for Video Transformers in Edge Camera Nodes

Recognizing human actions in video sequences has become an essential task in video surveillance applications. In such applications, transformer models have rapidly gained wide interest thanks to their performance. However, their advantages come at the cost of a high computational and memory cost, especially when they need to be incorporated in edge devices. In this work, temporal similarity tunnel insertion is utilized to reduce the overall computation burden in video transformer networks in action recognition tasks. Furthermore, an edge-friendly video transformer model is proposed based on temporal similarity, which substantially reduces the computation cost. Its smaller variant EMViT achieves 38% computation reduction under the UCF101 dataset, while keeping the accuracy degradation insignificant (<0.02%). Also, the larger variant CMViT reduces computation by 14% (13%) with an accuracy degradation of 2% (3%) in scaled Kinetic400 and Jester datasets.