Fourier-based adaptive counterfactual intervention for object re-identification.

Abstract

Object Re-Identification (ReID) is a fundamental task in computer vision, seeking to retrieve specific objects across diverse images. However, some non-identity cues (also known as confounders), such as background and color, often entangle discriminative identity representations, making it challenging to disentangle them in the spatial domain. Therefore, we present a novel method named Fourier-Based Adaptive Counterfactual Intervention (FACI-ReID) for object ReID. Specifically, FACI-ReID comprises two key components: the Counterfactual Intervention Module (CIM) and Fourier-based Selective Attention (FSA). First, two FSA modules extract factual and counterfactual features based on the Fourier transform, adaptively fusing frequency-domain information to capture sample-wise properties. Subsequently, CIM reduces the influence of confounders by maximizing the likelihood difference between factual and counterfactual features, thereby directing the network's attention to identity-intrinsic visual cues. Furthermore, counterfactual intervention in the frequency domain helps preserve identity-intrinsic information. With the above modules, our method learns more discriminative features for object ReID. Extensive experiments on five ReID benchmarks demonstrate that our proposed FACI-ReID achieves superior performance compared with state-of-the-art methods, proving its effectiveness.