Brief

Researchers have developed a new method called Counterfactual Intervention Feature Transfer (CIFT) to improve visible-infrared person re-identification. This technique addresses issues like modality imbalance between training and testing phases, and suboptimal graph topology learning. CIFT utilizes a Homogeneous and Heterogeneous Feature Transfer module to mitigate the modality gap and a Counterfactual Relation Intervention component to enhance the reliability of the graph topology structure. Experiments show CIFT surpasses existing state-of-the-art methods on standard benchmarks. AI

Researchers have introduced a new benchmark and dataset for Anytime Person Re-Identification (AT-ReID), designed to improve person retrieval across diverse conditions including varying times of day and long-term intervals. The AT-USTC dataset, collected over 21 months with 270 volunteers, offers a more extensive and varied dataset than previous efforts. To address the challenges of multi-scenario retrieval, a novel model called Uni-AT was proposed, featuring scenario-specific feature learning and strategies to manage inter-scenario interference and ensure balanced training. AI