Scholarlink

<p>Shoe image retrieval aims to recognize shoe models matching a target shoe’s appearance across various scenes and camera angles. This technique primarily supports the tracking of specific persons in urban surveillance system, with shoes used as the entry point. Traditional convolutional neural network (CNN)-based methods struggle with capturing global dependencies, while Vision Transformer (ViT) excels in this aspect but lack fine-grained local feature extraction. To address these challenges, we propose ViT-SIR, a shoe image retrieval network based on the transformer model. ViT-SIR integrates re-ranking, learnable prompt tuning, and advanced data augmentation methods. Additionally, we introduce a Channel Attention-Enhanced Multi-layer Perceptron (CAE-MLP) and a Local Feature Combination Learning Module (LFCM) to enhance both global and local feature extraction. Experimental results demonstrate that ViT-SIR achieves state-of-the-art performance on surveillance-sourced shoe image datasets, balancing accuracy, robustness, and training efficiency. Our approach significantly outperforms CNN-based and other transformer-based baseline models, highlighting its potential for real-world applications in criminal investigation and beyond. The code for the model proposed in this paper is available at: <a href="https://github.com/JisongSir/ViT_SIR">https://github.com/JisongSir/ViT_SIR</a>.</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

ViT-SIR: vision transformer-based shoe image retrieval with enhanced feature representation

  • 期刊论文

  • 发表日期：

    2025年6月10日

查看全文
  • Jisong Li,
  • Yunqi Tang

摘要

Shoe image retrieval aims to recognize shoe models matching a target shoe’s appearance across various scenes and camera angles. This technique primarily supports the tracking of specific persons in urban surveillance system, with shoes used as the entry point. Traditional convolutional neural network (CNN)-based methods struggle with capturing global dependencies, while Vision Transformer (ViT) excels in this aspect but lack fine-grained local feature extraction. To address these challenges, we propose ViT-SIR, a shoe image retrieval network based on the transformer model. ViT-SIR integrates re-ranking, learnable prompt tuning, and advanced data augmentation methods. Additionally, we introduce a Channel Attention-Enhanced Multi-layer Perceptron (CAE-MLP) and a Local Feature Combination Learning Module (LFCM) to enhance both global and local feature extraction. Experimental results demonstrate that ViT-SIR achieves state-of-the-art performance on surveillance-sourced shoe image datasets, balancing accuracy, robustness, and training efficiency. Our approach significantly outperforms CNN-based and other transformer-based baseline models, highlighting its potential for real-world applications in criminal investigation and beyond. The code for the model proposed in this paper is available at: https://github.com/JisongSir/ViT_SIR.