Kunming University of Science and Technology
The National Natural Science Foundation of China (General Program, Key Program, Major Research Plan)
The sketch person re-identification requires to search for pedestrians with the same identity as the given sketch image in the color image gallery. Due to the difference of posture and viewpoint between the sketch image and the color image, the two images from two different modes often have different semantic information in the same spatial position, which leads to the lack of robustness of the extracted features. Previous studies focused on pedestrian feature extraction modal-invariant information, but ignored the issue of semantic misalignment between different modal images, which led to the interference of the features by camera viewpoint, human posture or occlusion, which was not good for image matching. In this paper, a sketch pedestrian re-identification model based on channel information alignment is proposed, in which the semantic information alignment learning module guides the network to coded semantic information on the same channel of the feature to reduce the impact of misalignment of semantic information. Among them, the variant feature attention module assists the network to encode the variant identity related information, and designs the spatial variant regularization term to prevent the network from only paying attention to local features. The two modules cooperate with each other to strengthen the network"s perception and alignment of semantic information. Rank1and mAP of the proposed method in the challenging data set Sketch Re-ID and QMUL-ShoeV2 reach 60.0% and 59.3%, 33.5% and 46.1%, respectively, which verifies the effectiveness of the method.