Dual Gated Learning for Visible-Infrared Person Re-identification.

Visible-infrared person re-identification (VI-ReID) is a crucial part of open world ReID task, targeting at cross-modality pedestrian retrieval between visible and infrared images. Its large intra-class variation and cross-modality discrepancy lead to difficulty in discriminated representation learning. In this paper, we investigate adaptive neurons deactivation technique to improve VI-ReID model performance and propose two auxiliary training schemes. First, we present a one-stream module (gated module, GM) and its corresponding training scheme (gated learning, GL), to assist model training by adaptive neuron deactivation. Based on GM and GL, we design two-stream module (dual gated module, DGM) and its corresponding training scheme (dual gated learning, DGL) for further utilizing deactivated neurons in GL. During inference, GL and DGL are abandoned, resulting in no extra computation cost. Extensive experiments are performed on SYSU-MM01 and RegDB dataset to demonstrate the superiority of GL and DGL approach. Experimental results show that our proposed methods achieve significant improvement.