Abstract
In pedestrian re-identification, retrieving occluded pedestrians remains a challenging problem. The current methods primarily utilize additional networks to provide body cues for distinguishing the visible parts of the body. However, the inevitable domain gap between the auxiliary models and the Re-ID datasets significantly increases the difficulty in obtaining effective and efficient models. To eliminate the need for additional pre-trained networks, a Transformer-based dual correlation feature enhancement network model is proposed. Specifically, this method designs a relation-based feature enhancement module that effectively compensates for the absence or inaccuracy of local features by modeling the relational information within pedestrian images. Additionally, a dual correlation fusion module is designed to adaptively generate feature weights, fusing global and local features with weighted summation. Finally, extensive experiments were conducted on both occluded and holistic datasets to demonstrate that the proposed model outperforms state-of-the-art methods. The proposed model achieved a Rank-1 accuracy of 72.2% on the Occluded-Duke dataset and 88.0% on the Partial-REID dataset. This proves the effectiveness of the proposed approach.