Multi-Path and Multi-Loss Network for Person Re-Identification

Abstract

In person re-identification (re-ID), most state-of-the-art models extract features by convolutional neural networks to do similarity comparison. Feature representation becomes the key task for person re-ID. However, the learned features are not good enough based on a single-path and single-loss network because the learned objective only achieves one of the multiple minima. To improve feature representation, we propose a multi-path and multi-loss network (MPMLN) and concatenate multi-path features to represent pedestrian. Subsequently, we design MPMLN based on ResNet-50 and construct an end-to-end architecture. The backbone of our proposed network shares the local parameters for multiple paths and multiple losses. It has fewer parameters than multiple independent networks. Experimental results show that our MPMLN achieves the state-of-the-art performance on the public Market1501, DukeMTMC-reID and CUHK03 person re-ID benchmarks.