SCFNet: A Spatial-Channel Features Network based on Heterocentric Sample Loss for Visible-Infrared Person Re-Identification

Peng Su, Rui Liu, Jing Dong, Pengfei Yi, Dongsheng Zhou; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 3552-3568

Abstract


Cross-modality person re-identification between visible and infrared images has become a research hotspot in the image retrieval field due to its potential application scenarios. Existing research usually designs loss functions around samples or sample centers, mainly focusing on reducing cross-modality discrepancy and intra-modality variations. However, the sample-based loss function is susceptible to outliers, and the center-based loss function is not compact enough between features. To address the above issues, we propose a novel loss function called Heterocentric Sample Loss. It optimizes both the sample features and the center of the sample features in the batch. In addition, we also propose a network structure combining spatial and channel features and a random channel enhancement method, which improves feature discrimination and robustness to color changes. Finally, we conduct extensive experiments on the SYSU-MM01 and RegDB datasets to demonstrate the superiority of the proposed method.

Related Material


[pdf]
[bibtex]
@InProceedings{Su_2022_ACCV, author = {Su, Peng and Liu, Rui and Dong, Jing and Yi, Pengfei and Zhou, Dongsheng}, title = {SCFNet: A Spatial-Channel Features Network based on Heterocentric Sample Loss for Visible-Infrared Person Re-Identification}, booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)}, month = {December}, year = {2022}, pages = {3552-3568} }