Pose-to-Human (P2H): A pose guidance framework via Gram matrix for Occluded Person Re-identification

作     者：Quoc-Huy Trinh Phuoc-Thao Vo Thi Minh-Triet Tran Hai-Dang Nguyen 

作者机构：University of Science VNU-HCM Vietnam Software Engineering Laboratory University of Science VNU-HCM Vietnam 

出 版 物：《Procedia Computer Science》 

年 卷 期：2024年第246卷

页      面：1630-1639页

主　　题：Pose guidance Similarity matching Person Re-identification Gram Matrix 

摘      要：The person identification task aims to generate robust human representation embeddings. Traditional methods have achieved competitive results, but they face occlusion challenges, making it difficult to distinguish between humans occluded by objects or other humans. This paper proposes Pose-to-Human (P2H), a pose guidance framework via Gram matrix and knowledge transfer mechanism to tackle occlusion problems in Person Re-identification to address this issue. Our framework comprises four modules: Human Encoder, Pose Encoder, Compact Modules (CPM), and the Gram Matrix Guiding (GMG). The Pose Encoder generates pose information used to guide the human embedding models, facilitating learning of other parts of the human anatomy, which enables the model to focus on the unoccluded parts of the body, thus mitigating the reliance on the data for part-to-part matching, which is the first limitation of previous works. Additionally, it can reduce the demand for general appearance information from the human encoder model. Our method achieves competitive results through extensive experiments on five datasets compared to state-of-the-art methods, making it a promising framework for addressing the Re-Identification task.