Cross-modality Person Re-identification Method Using Cross-dimensional Interactive Attention Mechanism
Jin Wang, Kaiwei Jiang, Ze Lu, Xin Lu, Zhenbo She · 2023
The task of cross-modality person re-identification seeks to identify individuals from pedestrian images captured by diverse cameras, catering to the recognition needs across different scenarios. Variability in shooting time, location, and viewpoint introduces notable intra-modality and cross-modality discrepancies in the pedestrian images, diminishing recognition accuracy. Consequently, we proposed a cross-modality person re-identification technique (CDiNet), leveraging a cross-dimensional interactive attention mechanism. During the feature extraction phase, we integrated cross-dimensional interactive attention into the network, bolstering the feature representation. In the metric learning phase, we introduced a group loss, substituting the conventional triplet loss, and amalgamated it with identity loss to refine the network's constraints. The experimental results show that on the SYSU-MM01 dataset, the evaluation metrics Rank-1 and mAP achieve 54.43% and 52.46% respectively in All-search mode, and 62.73% and 68.96% in Indoor-search mode. The recognition accuracy of the proposed method CDiNet is better than similar methods.