Channel decoupling network for cross-modality person re-identification

Cross-modality person re-identification (CM-ReID) is a very challenging problem due to the discrepancy in data distributions between visible and near-infrared modalities. To obtain a robust sharing feature representation, existing methods mainly focus on image generation or feature constrain to decrease the modality discrepancy, which ignores the large gap between mixed-spectral visible images and single-spectral near-infrared images. In this paper, we address the problem by decoupling the mixed-spectral visible images into three single-spectral subspaces R, G, and B. By aligning the spectrum, we noted that even using a single spectral image instead of the VIS images could result in a better performance. Based on the above observation, we further introduce a clear and effective three-path channel decoupling network (CDNet) for combining the three spectral images. Extensive experiments implemented on the benchmark CM-ReID datasets, SYSU-MM01 and RegDB indicated that our method achieved state-of-the-art performance and outperformed existing approaches by a large margin. On the RegDB dataset, the absolute gain of our method in terms of rank-1 and mAP is well over 15.4% and 8.5%, respectively, compared with the state-of-the-art methods.