Heterogeneous information alignment and re-ranking for cross-modal pedestrian re-identification

Tiezhu Zhao, Xiaolun Liang, Kejing He, Qiuhong Yang, Ziliang Ren

Abstract: Cross-modal pedestrian re-identification, encompassing visible and infrared images, presents a significant challenge owing to inherent disparities in imaging principles, resulting in substantial cross-modal discrepancies. Consequently, matching and extracting heterogeneous information becomes exceedingly arduous. To address these issues, this research introduces a novel cross-modal pedestrian re-identification methodology, leveraging a heterogeneous information alignment algorithm and a reranking strategy. Specifically, a novel algorithm for aligning heterogeneous local information is proposed, which dynamically aligns the diverse local details of a given pedestrian by identifying the shortest path within the distance matrix of the pedestrian’s heterogeneous local information. Furthermore, to tackle the predicament of extracting heterogeneous pedestrian information, an expanded k-nearest neighbor reranking strategy is introduced, which integrates the heterogeneous information of the same pedestrian by dynamically developing the k-nearest neighbor heterogeneous information derived from the query image. Extensive experimental evaluations substantiate the effectiveness of the proposed approach across diverse scenarios.

External IDs:dblp:journals/mta/ZhaoLHYR25