SAMP: Sub-task Aware Model Pruning with Layer-Wise Channel Balancing for Person Search

Zimeng Wu, Jiaxin Chen, Yunhong Wang

Abstract: The deep convolutional neural network (CNN) has recently become the prevailing framework for person search. Nevertheless, these approaches suffer from the high computational cost, raising the necessity of compressing deep models for applicability on resource-restrained platforms. Despite of the promising performance achieved in boosting efficiency for general vision tasks, current model compression methods are not specifically designed for person search, thus leaving much room for improvement. In this paper, we make the first attempt in investigating model pruning for person search, and propose a novel loss-based channel pruning approach, namely Sub-task Aware Model Pruning with Layer-wise Channel Balancing (SAMP). It firstly develops a Sub-task aware Channel Importance (SaCI) estimation to deal with the inconsistent sub-tasks, i.e. person detection and re-identification, of person search. Subsequently, a Layer-wise Channel Balancing (LCB) mechanism is employed to progressively assign a minimal number of channels to be preserved for each layer, thus avoiding over-pruning. Finally, an Adaptive OIM (AdaOIM) loss is presented for pruning and post-training via dynamically refining the degraded class-wise prototype features by leveraging the ones from the full model. Experiments on CUHK-SYSU and PRW demonstrate the effectiveness of our method, by comparing with the state-of-the-art channel pruning approaches.