DART³: Test-time latent geometry adaptation for re-identification

Person re-identification (ReID) models deployed in multi-camera surveillance networks are vulnerable to camera bias. Learned embeddings often cluster by camera viewpoint and background rather than identity, leading to a degradation of performance when new cameras are added. This paper first provides an empirical analysis of this effect in SoTA ReID backbones and observes that camera-induced shifts behave approximately as a linear scale–shift distortion in the feature space. At the same time, we observe that the Euclidean distance between a query and its nearest true gallery neighbor provides a much more stable indicator of error than classifier entropy. Motivated by these findings, we introduce DART³ (Distance-Aware Retrieval Tuning at Test Time), a test-time adaptation method that directly shapes the query–gallery geometry under a linear-camera model. DART³ treats the pretrained ReID backbone as a black box and adapts only a small set of camera-specific affine parameters, trained to reduce top-k Euclidean distances to likely positives while pushing other gallery items away, without accessing source data or modifying backbone weights. Empirical evaluations on multiple person and vehicle ReID benchmarks show that DART³ consistently improves over entropy-based TTA and static camera-normalization baselines, while incurring only a small fraction of their inference-time overhead, making it a practical tool for mitigating camera bias in real-world deployments.