Scholarlink

<p>Visual transformer (ViT) has been successfully applied in computer vision and recently in person re-identification. However, the approach of feeding all image patches into the self-attention module can lead to computational redundancy and noise, as not all patches contain discriminative information. To address this issue, this paper proposes a Token Recombination based Shallow-Deep feature Fusion framework (TRSDF). Specifically, it contains two modules. (1) In order to eliminate the interference of occlusion and enrich the pedestrian feature representation, an attention matrix-based Token Recombination Module (TRAM) is proposed. First, we construct a token filtering mechanism that classifies all image tokens into irrelevant tokens and target tokens, according to the effective amount of information contained in the token. The input sequence is reselected by the filtering operation to reduce the attention to the background and occlusion. Then, by designing a token replacement strategy, the discriminative information in the target token is embedded into the irrelevant token, and the original feature is reconstructed, which further enriches the original feature representation. (2) To effectively improve the expressiveness of the model, a Shallow-Deep Feature-based learning Module (SDFM) is constructed that strengthens the correlation between the layers of the network. It treats the deep semantic feature branch as a query and exchanges information with the class token of low-level features in the network’s shallow layers to strengthen the robust feature representation through the network’s depth. Finally, we conduct extensive experiments on various public datasets and obtain competitive results through quantitative evaluation. Code is available at <a href="https://github.com/moonlight0427/TRSDF">https://github.com/moonlight0427/TRSDF</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

Token recombination based shallow-deep feature fusion for occluded person re-identification

  • 期刊论文

  • 发表日期：

    2025年6月6日

查看全文
  • Shuze Geng,
  • Yifan Liu,
  • Gang Yan,
  • Haowei Wang,
  • Pengfei Zhao,
  • Wenjie Xia

摘要

Visual transformer (ViT) has been successfully applied in computer vision and recently in person re-identification. However, the approach of feeding all image patches into the self-attention module can lead to computational redundancy and noise, as not all patches contain discriminative information. To address this issue, this paper proposes a Token Recombination based Shallow-Deep feature Fusion framework (TRSDF). Specifically, it contains two modules. (1) In order to eliminate the interference of occlusion and enrich the pedestrian feature representation, an attention matrix-based Token Recombination Module (TRAM) is proposed. First, we construct a token filtering mechanism that classifies all image tokens into irrelevant tokens and target tokens, according to the effective amount of information contained in the token. The input sequence is reselected by the filtering operation to reduce the attention to the background and occlusion. Then, by designing a token replacement strategy, the discriminative information in the target token is embedded into the irrelevant token, and the original feature is reconstructed, which further enriches the original feature representation. (2) To effectively improve the expressiveness of the model, a Shallow-Deep Feature-based learning Module (SDFM) is constructed that strengthens the correlation between the layers of the network. It treats the deep semantic feature branch as a query and exchanges information with the class token of low-level features in the network’s shallow layers to strengthen the robust feature representation through the network’s depth. Finally, we conduct extensive experiments on various public datasets and obtain competitive results through quantitative evaluation. Code is available at https://github.com/moonlight0427/TRSDF.