| Title |
Transformer-based Cross attention and Feature Diversity for Occluded Person Re-identification |
| Authors |
강성재(Sungjae Kang) ; 김세준(Sejun Kim) ; 서기성(Kisung Seo) |
| DOI |
https://doi.org/10.5370/KIEE.2023.72.1.108 |
| Keywords |
Deep learning; Occluded Person Re-ID; Transformer; keypoint Heatmap; Cross-attention; Feature Diversity |
| Abstract |
Occluded person re-identification is a very difficult because the specific person is occluded by obstacles or other persons or by oneself. Major works adopt transformer-based approach show excellent performances, but they used a basic transformer only. In this paper, we suggest the various techniques to improve the transformer-based Re-ID method for the occluded person as follows. First, after extracting the heatmap and then deleting random body parts on the heat map, accurate keypoint information is obtained in data augmentation. Second, Cross-attention between the keypoint heatmap and the output of the transformer's middle layer is provided to focus more on the non-occluded person area. Third K-menas clustering is utilized to enhance the representation of local features, and the structure of the network is proposed to improve the diversity of the features. We evaluate mAP and Rank-1 performance on the Occluded-Duke and Market-1501 dataset and compare the proposed model with existing state-of-the-art techniques. Experimental results show that our method outperforms state-of-the-art methods. |