Diversified Distillation Fusion Network for vehicle re-identification

Abstract

Vehicle re-identification (Re-ID), which focuses on matching vehicles through numerous cameras with non-overlapping views, has significant applications in intelligent transportation systems. A critical challenge for vehicle Re-ID is to extract abundant and discriminative features from vehicle images while effectively connecting global and local features for optimal complements. To address this challenge, we propose a novel vehicle Re-ID model that not only fuses multi-level features of the backbone network but also acquires global and local features without employing extra detection or segmentation networks. We first introduce Cross Spatial and Channel Fusion (CSCF) module to integrate spatial information from low-level feature maps and semantic information from high-level feature maps such that complementary features from different levels can be leveraged to enhance identification performance. The channel partition is then applied to generate multiple local features without requiring an additional segmentation network. To reveal the important local features, we further design Deep Relation Spatial Attention (DRSA) module that reinforces the connections between different regions of the local feature maps. Furthermore, we propose Diversity Regularized Self-Distillation (DRSD), where the global representation guides local branches to learn distinct and complementary knowledge. This innovative design enhances the specialization and improves the generalization and performance of the model. Experimental results on multiple large-scale vehicle Re-ID benchmark datasets demonstrate that our proposed Diversified Distillation Fusion Network (DDFN) reaches state-of-the-art performance. Furthermore, ablation studies are conducted with abundant detail to prove the effectiveness of our model design.