PaTTA-ID : Practical Test-Time Adaptation for Person Re-Identification

ICLR 2026 Conference Submission8828 Authors

Keywords: person re-identification, test-time adaptation, image retrieval, computer vision

Abstract:

Existing test-time adaptation (TTA) methods for person re-identification (re-ID) assume unrealistic scenarios: a large target gallery is available in advance, ignores temporal correlations in streaming input, and all identities are guaranteed to exist in the gallery set. Furthermore, they rely on server-side settings where data from multiple cameras are aggregated in advance, which is unrealistic for edge
device applications on a single-camera. Therefore, they experience performance degradation in practical real-world deployments due to domain gaps between the training (source) data and the unseen (target) gallery streams. In this work, we introduce a practical scenario of test-time adaptation for person re-ID tailored for online streaming environment on resource-constrained edge devices, where a
small predefined query set is registered in advance and unlabeled large gallery data continuously arrive from a single camera stream. We propose a novel framework to address this practical problem, called PaTTA-ID, that enables effective adaptation from two complementary perspectives. First we devise Input Distribution Compensation, which employs query-guided sampling and contrastive adaptation
to compensate the bias of streaming inputs and promote cross-camera discriminability. Moreover, we investigate Model Drift Compensation, which prevents the bias toward the current camera stream via camera invariant learning and query
feature compensation. Experimental results evaluated on four benchmark datasets compared with nine baselines demonstrate that the proposed PaTTA-ID achieves state-of-the-art performance surpassing existing TTA methods.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 8828