DisenCB: Learning Diverse Disentangled Representation for Cross-Domain Consecutive Footprint Retrieval

Cross-domain consecutive footprint retrieval is crucial for security screening in the security fields, which faces the challenges of large intra-class differences and mining of temporal information. This paper propose a cross-domain consecutive footprint retrieval network based on diverse disentangled representation, mining more discriminative footprint features and domain-invariant temporal information. First, we separate the footprint features into domain-relevant and domain-irrelevant features by the disentangled module; second, we introduce a complementary attention fusion module that effectively interacts domain-relevant and domain-irrelevant features to obtain a more robust feature representation; finally, the bi-directional temporal enhancement module makes full use of domain-irrelevant features while taking into account both forward and backward frame information. Experiments show that the network proposed in this paper obtains 77.78% Rank-1 on the CDSPF-200 dataset, which outperforms other retrieval networks.