Visible Infrared Person Re-Identification via Global-Level and Local-Level Constraints

Visible infrared person re-identification (VI-ReID) is an extremely challenging task. VI-ReID suffers from two challenges. One is the cross-modality discrepancy due to different camera spectrums, the other is the intra-modality variation caused by the noise of background clutter or occlusion. We propose a global-level and local-level constraints network (GLoC-Net) to learn discriminative feature representations. It mainly contains two aspects. 1) We employ a non-local attention mechanism for extracting shared features to mitigate the cross-modality discrepancy, and present the division operation of local features to alleviate the problem that the non-local attention mechanism is less robust to noise. 2) We propose joint constraints of global-level and local-level to alleviate the intra-modality variation, which makes the algorithm more robust to noise. Experiments demonstrate that the superior performance of proposed method compared with the state-of-the-arts.