Exploring the Camera Bias of Person Re-identification

TL;DR

This work investigates camera bias in person ReID, showing that unseen-domain shifts amplify bias and that unsupervised models exhibit bias even in seen domains. The authors introduce camera-specific feature normalization as a simple, test-time postprocessing step—computing per-camera means and standard deviations and applying $\\hat{f}_i = (\\boldsymbol{f}_i - \\boldsymbol{m}_{\\mathbf{y}_i}) / \\boldsymbol{\\sigma}_{\\mathbf{y}_i}$—and demonstrate its effectiveness across diverse models and benchmarks. Through detailed analysis of feature-space structure, low-level image properties, and body-angle factors, they explain why normalization reduces bias and show how it generalizes beyond camera labels to detail-level biases. They also address the risk of camera bias in unsupervised learning, proposing simple, effective strategies (debiased pseudo labeling and discarding biased clusters) that yield substantial improvements, thereby offering practical, scalable solutions for robust ReID in unseen domains.

Abstract

We empirically investigate the camera bias of person re-identification (ReID) models. Previously, camera-aware methods have been proposed to address this issue, but they are largely confined to training domains of the models. We measure the camera bias of ReID models on unseen domains and reveal that camera bias becomes more pronounced under data distribution shifts. As a debiasing method for unseen domain data, we revisit feature normalization on embedding vectors. While the normalization has been used as a straightforward solution, its underlying causes and broader applicability remain unexplored. We analyze why this simple method is effective at reducing bias and show that it can be applied to detailed bias factors such as low-level image properties and body angle. Furthermore, we validate its generalizability across various models and benchmarks, highlighting its potential as a simple yet effective test-time postprocessing method for ReID. In addition, we explore the inherent risk of camera bias in unsupervised learning of ReID models. The unsupervised models remain highly biased towards camera labels even for seen domain data, indicating substantial room for improvement. Based on observations of the negative impact of camera-biased pseudo labels on training, we suggest simple training strategies to mitigate the bias. By applying these strategies to existing unsupervised learning algorithms, we show that significant performance improvements can be achieved with minor modifications.

Figures (15)

• Figure 1: Cosine distance distributions of a camera-aware ReID model on (a) the training domain (Market-1501) and (b) the unseen domain (MSMT17). The distances between samples within the same cameras are more skewed to the left when the data distribution is shifted.
• Figure 2: Analysis on the 384-dimensional embedding space of a ReID model. We measure the similarity of displacement vectors and mAP results increasing the number of feature dimensions following different orders. (a) Variance of each dimension of camera mean features. (b) Cosine similarity of displacement vectors between samples of the same identities from different cameras along selected dimensions. (c) Result of camera-specific feature centering for selected dimensions.
• Figure 3: Analysis on low-level properties. (a) Cosine similarity of displacement vectors by image transformations. (b) Property group-specific feature normalization. The dashed line indicates the performance without normalization. (c) (Property group, camera)-specific feature normalization. The dashed line indicates the performance with camera-specific (and property-agnostic) normalization.
• Figure 4: Normalization result of Figure \ref{['fig:motivation']}(b).
• Figure 5: Results based on the number of samples used to calculate normalization parameters.