Title: Advancing Towards Anytime Retrieval: A New Benchmark for Anytime Person Re-Identification

Abstract:

In practical applications, person re-identification (ReID) systems are required to locate target individuals at any given moment, encompassing both day and night conditions and spanning from short-term to long-term intervals. However, current ReID tasks and datasets fall short of this demand, as they are limited by specific timeframes and only offer training and evaluation for isolated scenarios. To bridge this gap, we introduce a novel task termed Anytime Person Re-identification (AT-ReID), designed to facilitate effective retrieval across diverse temporal scenarios.

To support this new challenge, we have compiled AT-USTC, the inaugural large-scale dataset for AT-ReID. This collection comprises 403,000 images of individuals in various attire, captured using both RGB and infrared (IR) cameras. The data gathering process extended over 21 months, involving 270 volunteers who were photographed an average of 29.1 times across different dates and locations. This frequency is four to fifteen times higher than that found in existing datasets, thereby establishing a robust foundation for future research in AT-ReID.

Furthermore, we propose Uni-AT, a unified model designed to address the complexities of multi-scenario retrieval. Uni-AT integrates a multi-scenario ReID (MS-ReID) framework for learning scenario-specific features, a Mixture-of-Attribute-Experts (MoAE) module to mitigate interference between scenarios, and a Hierarchical Dynamic Weighting (HDW) strategy to maintain balanced training across all contexts. Extensive experimental results demonstrate that our approach yields satisfactory performance and exhibits strong generalization capabilities across all tested scenarios.

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