Title: Counterfactual Intervention Feature Transfer for Visible-Infrared Person Re-identification

Abstract:

Graph-based approaches have recently demonstrated significant success in person re-identification (ReID) by first calculating the graph topology (affinities) among individuals and then propagating information across them to generate robust features. However, we identify that existing graph-based methods for visible-infrared person re-identification (VI-ReID) suffer from poor generalization due to two primary factors. First is the train-test modality balance gap, an inherent characteristic of the VI-ReID task; while data from both modalities are balanced during training, they become highly unbalanced during inference, leading to reduced generalization performance. Second is the sub-optimal topology structure resulting from the end-to-end learning approach applied to the graph module. We attribute this to the fact that well-trained input features hinder the learning of the graph topology, preventing it from generalizing effectively during inference.

To address these challenges, this paper introduces Counterfactual Intervention Feature Transfer (CIFT). Specifically, we design a Homogeneous and Heterogeneous Feature Transfer (H2FT) mechanism to bridge the train-test modality balance gap. This is achieved through two distinct, carefully crafted graph modules alongside a simulation of unbalanced scenarios. Additionally, we propose Counterfactual Relation Intervention (CRI), which leverages counterfactual intervention and causal effect techniques to emphasize the importance of the topology structure throughout the training process, thereby enhancing the reliability of the graph topology. Extensive experiments conducted on standard VI-ReID benchmarks confirm that CIFT surpasses state-of-the-art methods across various settings.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC