Mask‐guided class activation mapping network for person re‐identification

Abstract

In this Letter, the authors propose a novel mask‐guided class activation mapping (MCAM) network for person re‐identification, which learns background‐invariant and view‐invariant features. Specifically, a novel loss function named mask‐guided mapping loss is meticulously formulated to utilise the human binary masks, which contain helpful body shape information as the reference standard, thereby guiding the model to place more emphasis on human body regions. Moreover, they propose a new weighted channel attention (WCA) module, which replaces the global average pooling with a global depthwise convolution layer. By virtue of this particular WCA module, the feature information distributed across the spatial space can be individually weighted and dynamically compressed into a more precise channel attention map. Extensive experiments have been carried out on three widely‐used re‐identification data sets. Compared with the baseline model, MCAM has gained rank‐1 accuracy improvement of 2.0% on Market‐1501, 6.0% on DukeMTMC‐reID, and 7.5% on CUHK03‐NP, confirming its effectiveness.