Scholarlink

<p>Person Re-identification (ReID) aims to generate a discriminative description model to search the probe person from the gallery images. Previous methods infer the ReID model by constructing the metric learning between the visual space and the annotated label space. Moreover, the textual knowledge inferred by the visual-language model is introduced in CLIP-ReID to enhance the descriptive ability of the ReID model. However, the textual knowledge inferred from the pre-trained visual space has less discriminative ability on ReID tasks. To address the above issue, we propose a novel Visual-Textual Adversarial Learning (VTAL) for person ReID. The primary concept of VTAL is to construct an adversarial loop between the visual encoder and the text encoder, leveraging the progressive enhancement of one encoder to improve the performance of the other within this loop. Two types of prompts (Task-Independent prompt and Task-Related prompt) are deployed to maintain the generalization ability and discrimination ability of the generated textual-level identity embedding simultaneously. After that, the generated corresponding identity embeddings are treated as a textual-to-visual constraint to optimize the visual encoder. Extensive experiments on three benchmarks verify the effectiveness of the proposed method for person ReID.</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

Visual-textual adversarial learning for person re-identification

  • 期刊论文

  • 发表日期：

    2025年1月16日

查看全文
  • Pengqi Yin

摘要

Person Re-identification (ReID) aims to generate a discriminative description model to search the probe person from the gallery images. Previous methods infer the ReID model by constructing the metric learning between the visual space and the annotated label space. Moreover, the textual knowledge inferred by the visual-language model is introduced in CLIP-ReID to enhance the descriptive ability of the ReID model. However, the textual knowledge inferred from the pre-trained visual space has less discriminative ability on ReID tasks. To address the above issue, we propose a novel Visual-Textual Adversarial Learning (VTAL) for person ReID. The primary concept of VTAL is to construct an adversarial loop between the visual encoder and the text encoder, leveraging the progressive enhancement of one encoder to improve the performance of the other within this loop. Two types of prompts (Task-Independent prompt and Task-Related prompt) are deployed to maintain the generalization ability and discrimination ability of the generated textual-level identity embedding simultaneously. After that, the generated corresponding identity embeddings are treated as a textual-to-visual constraint to optimize the visual encoder. Extensive experiments on three benchmarks verify the effectiveness of the proposed method for person ReID.