Optimal Topology of Vision Transformer for Real-Time Video Action Recognition in an End-To-End Cloud Solution

This study introduces an optimal topology of vision transformers for real-time video action recognition in a cloud-based solution. Although model performance is a key criterion for real-time video analysis use cases, inference latency plays a more crucial role in adopting such technology in real-world scenarios. Our objective is to reduce the inference latency of the solution while admissibly maintaining the vision transformer's performance. Thus, we employed the optimal cloud components as the foundation of our machine learning pipeline and optimized the topology of vision transformers. We utilized UCF101, including more than one million action recognition video clips. The modeling pipeline consists of a preprocessing module to extract frames from video clips, training two-dimensional (2D) vision transformer models, and deep learning baselines. The pipeline also includes a postprocessing step to aggregate the frame-level predictions to generate the video-level predictions at inference. The results demonstrate that our optimal vision transformer model with an input dimension of 56 x 56 x 3 with eight attention heads produces an F1 score of 91.497% for the testing set. The optimized vision transformer reduces the inference latency by 40.70%, measured through a batch-processing approach, with a 55.63% faster training time than the baseline. Lastly, we developed an enhanced skip-frame approach to improve the inference latency by finding an optimal ratio of frames for prediction at inference, where we could further reduce the inference latency by 57.15%. This study reveals that the vision transformer model is highly optimizable for inference latency while maintaining the model performance.