Multi-scale feature correspondence and restriction mechanism for visible X-ray baggage re-Identification

Sixian Chan, Jiaao Cui, Yonggan Wu, Hongqiang Wang, Cong Bai

Abstract: Recently, social security surveillance has posed a new AI challenge, i.e., Visible-X-ray baggage Re-Identification (VX-ReID), which aims to re-identify and retrieve baggage between visible and X-ray imaging modalities. Compared with cross-modality person re-identification, VX-ReID has two distinctive bottlenecks: shape deformation and feature entanglement. For the former, the shape of the baggage can change largely, resulting in serious feature unrobustness. For the latter, the X-ray images often contain the contents of the baggage, which are not visible in daylight images. These will greatly affect the performance of representational learning loss functions (like ID Loss) in the Re-ID task. In this paper, we propose a cross-modality multi-scale feature correspondence model (CMMFC) for VX-ReID. Specifically, we devise and calculate multiple feature correspondences between modalities on multiple-scale feature maps endowed to overcome the deformation problem. We also utilize a novel feature restriction mechanism (FRM) to alleviate the feature entanglement problem, which imposes different constraints on features at different scales and accurately drives networks to distinctive modality-irrelevant features. Finally, CMMFC is extensively evaluated on our dataset RX01. Experiments show that our proposed method achieves state-of-the-art performance on dataset RX01.