SAM-driven MAE Pre-training and Background-aware Meta-learning for Unsupervised Vehicle Re-identification

Abstract

This paper addresses background interference in unsupervised vehicle re-identification by combining SAM-driven masked autoencoder pre-training with background-aware meta-learning. The method first uses SAM to separate vehicle identity regions from background regions, and further introduces a spatially constrained background segmentation strategy to handle difficult cases such as occlusion and ambiguous vehicle boundaries. Based on the optimized segmentation results, SAM-driven MAE pre-training selectively preserves vehicle-related patches and masks background regions, encouraging the encoder to learn identity-sensitive representations in a self-supervised manner. To improve robustness under varying scene conditions, the paper further proposes a background-aware meta-learning strategy that constructs meta-training and meta-testing splits according to different background region ratios. Experiments on VeRi-776 and VeRi-Wild demonstrate that the proposed method effectively reduces background interference and improves unsupervised vehicle Re-ID performance.