Fusion Shape-Guided Deletion for person Re-identification

In the field of infrared-visible pedestrian re-identification (Visual-Infrared Reid), the significant differences between visible light and infrared modalities pose a significant challenge in re-identifying and matching individuals across different scenes and viewpoints. Given the limited training samples and the vast differences between modalities, this paper proposes a novel enhancement network, the Fusion Shape-Guided Deletion (FSGD-Reid) network. Through FSGD, the network can not only emphasize and enhance the features related to shape within the same matching feature region, but also adaptively allow the preservation of some relevant features outside the shape, ensuring that the features for cross-modal re-identification are retained more completely without compromising the shape deletion effect. In summary, first, FSGD adaptively fuses the original image with the shape image to highlight shape-related features, enabling incomplete segmentation and potential loss of some features outside the matching feature region but still relevant for cross-modal re-identification. Then, to suppress the redundant and useless features resulting from this, the FSGD network also performs deeper feature enhancement and modality fusion. Finally, for relatively ambiguous re-identification scenarios, this paper attempts a hybrid approach for shape segmentation and fusion. Ultimately, compared to multiple state-of-the-art (SOTA) methods, the proposed approach achieves leading results on relevant metrics.