‹ back to 2026-07-16

Structural-Semantic Reciprocal Learning for Unsupervised Visible-Infrared Person Re-Identification

arXiv cs.CV Multimodal & Generative Moyao Tian, Shijia Liu, Yan Yang, Xin Yuan, Minshi Chen, Wei Wang, Xiao Wang 2026-07-16

TL;DR - A new framework (SSRL) tackles unsupervised visible-infrared person re-identification by turning noisy open-loop cross-modal association into a self-correcting closed-loop system, improving robustness without identity annotations.

Problem: USVI-ReID struggles with a large modality gap and no cross-modal labels; prior progressive-association methods rely on ambiguous global features and propagate pseudo-label noise unchecked.
Fine-grained Structural Decoupling (FSD): Extracts discriminative body-part primitives as reliable spatial anchors to complement ambiguous holistic silhouettes.
Closed-loop Semantic Calibration (CSC): Rebuilds shared semantic prototypes each epoch and feeds them back to filter pseudo-label noise before the next clustering cycle.
Results: Competitive with state-of-the-art USVI-ReID methods on SYSU-MM01 and RegDB, reportedly surpassing several supervised approaches on RegDB (no specific metrics given in the provided content).

view merged work →