Scholarlink

<p>Vehicle re-identification (Re-ID) is pivotal for intelligent transportation systems but remains challenging under severe occlusions and viewpoint variations, where existing models often fail by either incorporating distracting background noise or losing critical identity (ID)-specific information. To address the limitations of current datasets, restricted viewpoints, limited scene diversity, and insufficient occlusion samples, this paper introduces two new benchmarks: <span>Vehicle-Syn</span>, a large-scale synthetic dataset generated via game engines to simulate diverse weather and illumination conditions, and <span>Occluded-Wild</span>, a real-world dataset constructed upon <span>VeRI-Wild</span> to specifically evaluate occlusion robustness. Motivated by the need for robust feature purification and complementation, we propose a novel Occlusion-Aware Re-ID framework (OARI) that operates on a progressive logic of filtering, retrieval, and fusion. Our framework employs a sparse feature encoder to dynamically discard non-essential image tokens related to occlusions, followed by a feature matching module that retrieves <i>K</i> neighboring gallery images using joint image-level and patch-level similarity. Finally, a Transformer-based fusion module integrates complementary information from these neighbors to compensate for occlusion-induced feature loss. Experimental results demonstrate the effectiveness of our approach, achieving 75.2% Rank-1 accuracy on <span>Occluded-Wild</span> and 80.1% on <span>Vehicle-Syn</span>, yielding competitive performance compared to state-of-the-art methods under severe occlusion.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Occlusion-aware vehicle re-identification via sparse token modeling under multi-perspective cues

  • 期刊论文

  • 发表日期：

    2026年7月19日

查看全文
  • Yuanbang Li,
  • Xianyao Ping,
  • Rong Peng,
  • Yichen Liu,
  • Xian Zhong

摘要

Vehicle re-identification (Re-ID) is pivotal for intelligent transportation systems but remains challenging under severe occlusions and viewpoint variations, where existing models often fail by either incorporating distracting background noise or losing critical identity (ID)-specific information. To address the limitations of current datasets, restricted viewpoints, limited scene diversity, and insufficient occlusion samples, this paper introduces two new benchmarks: Vehicle-Syn, a large-scale synthetic dataset generated via game engines to simulate diverse weather and illumination conditions, and Occluded-Wild, a real-world dataset constructed upon VeRI-Wild to specifically evaluate occlusion robustness. Motivated by the need for robust feature purification and complementation, we propose a novel Occlusion-Aware Re-ID framework (OARI) that operates on a progressive logic of filtering, retrieval, and fusion. Our framework employs a sparse feature encoder to dynamically discard non-essential image tokens related to occlusions, followed by a feature matching module that retrieves K neighboring gallery images using joint image-level and patch-level similarity. Finally, a Transformer-based fusion module integrates complementary information from these neighbors to compensate for occlusion-induced feature loss. Experimental results demonstrate the effectiveness of our approach, achieving 75.2% Rank-1 accuracy on Occluded-Wild and 80.1% on Vehicle-Syn, yielding competitive performance compared to state-of-the-art methods under severe occlusion.