Dynamic Channel Pruning for Real-Time Object Detection Networks

摘要

Real-time object detection is a highly practical technical means which has been the focus of research in recent years. While the high requirement of running memory and computing resources hinder the deployment on resource-limited devices. In this paper, we propose an effective method to dynamically enhance the sparsity on channel-level. To this end, we introduce dynamic sparsity coefficient (DSC) to balance model training and sparse training as well as adaptable sparsity regularization (TLp) to reinforce sparsity. We monitor the saliency of channels to evaluate their significance for model performance during the sparse training process, according to which, we implement different sparsity strategy on channels. Aiming to maintain the representation ability of important parameters on a fine-grained level, we automatically discern insignificant channels and remove with channel pruning. We demonstrate our method on latest object detector YOLOv4 and lightweight model YOLOv4-Tiny. Compare with uncompressed model, our method can obtain 85.8% decrease of FLOPs, 88.9% declines of parameters with a moderate accuracy loss. Compared with other model compression methods, we achieves comparable results with fewer trainable parameters but better detection performance.