Scholarlink

<p>This paper introduces a method based on the transformer-driven TransReID architecture, which employs the Vision Transformer (ViT) framework to address significant Person Re-Identification (Re-ID) challenges, including occlusion, viewpoint changes, and feature misalignment. The primary contribution is the proposed Occlusion Augmentation with Extensive Variation (OAVE) module, which augments the training dataset by introducing realistic occlusions. This process trains the model to focus on visible, distinctive features while disregarding occluded regions. In addition to OAVE, the framework incorporates two existing modules: the Jigsaw Patch Module (JPM) and Side Information Embeddings (SIE), which further enhance model robustness. JPM improves feature learning by rearranging image patches, while SIE encodes non-visual information to make the framework resilient to camera biases. A re-ranking post-processing technique refines retrieval results using mutual nearest-neighbor relationships. Extensive experiments demonstrate the effectiveness of the proposed architecture. The method achieved mAP and Rank-1 accuracies of 74.9% and 76.2% on the occlusion-specific Occluded Duke dataset and 87.2% and 90.3% on the P-Duke dataset, respectively. Furthermore, it sets a new benchmark of 91.3% mAP and 97.2% Rank-1 accuracy on the holistic Market1501 dataset, substantially outperforming state-of-the-art methods. These results underline the improved accuracy and robustness under both occluded and non-occluded conditions.</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

Robust vision transformer-based framework for person re-identification through occlusion-aware training

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • Mohammed Abusarie Fouad,
  • Hanaa M. Hamza,
  • Khalid M. Hosny

摘要

This paper introduces a method based on the transformer-driven TransReID architecture, which employs the Vision Transformer (ViT) framework to address significant Person Re-Identification (Re-ID) challenges, including occlusion, viewpoint changes, and feature misalignment. The primary contribution is the proposed Occlusion Augmentation with Extensive Variation (OAVE) module, which augments the training dataset by introducing realistic occlusions. This process trains the model to focus on visible, distinctive features while disregarding occluded regions. In addition to OAVE, the framework incorporates two existing modules: the Jigsaw Patch Module (JPM) and Side Information Embeddings (SIE), which further enhance model robustness. JPM improves feature learning by rearranging image patches, while SIE encodes non-visual information to make the framework resilient to camera biases. A re-ranking post-processing technique refines retrieval results using mutual nearest-neighbor relationships. Extensive experiments demonstrate the effectiveness of the proposed architecture. The method achieved mAP and Rank-1 accuracies of 74.9% and 76.2% on the occlusion-specific Occluded Duke dataset and 87.2% and 90.3% on the P-Duke dataset, respectively. Furthermore, it sets a new benchmark of 91.3% mAP and 97.2% Rank-1 accuracy on the holistic Market1501 dataset, substantially outperforming state-of-the-art methods. These results underline the improved accuracy and robustness under both occluded and non-occluded conditions.