Cross-modal person re-identification using fused local effective features and multi-scale features

Abstract

The main research objective of cross-modal person re-identification is to retrieve matching images of the same person from image repositories in both modalities, given visible light or infrared images of individuals. Due to the significant modality gap between pedestrian images, the task of person re-identification faces considerable challenges. To address this issue, a method is proposed that utilizes the fusion of local effective features and multi-scale features. First, images are transformed into pseudo-infrared images through data augmentation and then a dual-stream network is designed using ResNet50_IBN for feature extraction. Subsequently, pedestrian features extracted from different layers are fused at multiple scales to alleviate feature loss caused during the convolution process. Finally, the model is supervised using global features and local effective features to address issues related to cluttered backgrounds and varying pedestrian positions in images. The proposed method is experimentally validated on the current mainstream cross-modal person re-identification datasets SYSU-MM01 and RegDB, demonstrating improvements in Rank-1 and mAP metrics compared to current state-of-the-art algorithms.