Vehicle Re-Identification in Occluded Scenes Based on Vision Transformer

Occlusion introduces significant challenges to vehicle re-identification. To solve this problem, we proposes an occluded vehicle re-identification method (SMR) based on sparse feature encoder and feature recovery module. The proposed method uses a sparse feature encoder to discard less informative image tokens by leveraging the relevance in classification token attention. Subsequently, the feature matching module identifies the top-K nearest gallery images using the retained tokens through a combination of image-level and patch-level similarity. Finally, the feature recovery module compensates for the pruned features by incorporating information from the nearest gallery images. Besides, we construct a large-scale dataset for the Occluded Vehicle Re-ID, namely Occluded-Wild, which is by far the largest dataset for the Occlusion Vehicle Re-ID. Extensive experiments are conducted on our constructed occluded re-id dataset and a commonly used holistic re-id dataset. Our method largely outperforms existing vehicle re-id methods on these two datasets.