Scholarlink

<p>Accurate human pose estimation remains challenging, particularly in complex occlusion and truncation scenarios. Traditional methods often struggle with the ambiguities introduced by occluded poses. To address this issue, we propose a novel progressive dual-branch transformer-based diffusion model that explicitly considers pose uncertainty. Our model integrates conventional pose estimation and segmentation methods to initialize distributions, which accelerate the denoising process. A transformer with a cross-attention module is introduced to facilitate information sharing between the pose and mask branches, leading to improved refinement of noisy poses. Experimental results on the MS COCO, 3DMPB, and CrowdPose datasets demonstrate that our approach outperforms baseline methods, achieving a significant increase in average precision (AP). Our dual-branch transformer-based diffusion model not only reduces pose uncertainty but also enhances the efficiency of the reverse diffusion process by utilizing initial noise distributions derived from initial heatmaps and masks. This work highlights the potential of diffusion models in addressing the challenges of 2D human pose estimation under occlusion. The code is available at <a href="https://github.com/wenhappyall/ProDiffPose">https://github.com/wenhappyall/ProDiffPose</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

Progressive dual-branch transformer-based diffusion model: a novel approach for robust 2D human pose estimation

  • 期刊论文

  • 发表日期：

    2025年7月7日

查看全文
  • Huilin Liu,
  • Xinyue Wen,
  • Xinwei Ye,
  • Wenkang Zhang

摘要

Accurate human pose estimation remains challenging, particularly in complex occlusion and truncation scenarios. Traditional methods often struggle with the ambiguities introduced by occluded poses. To address this issue, we propose a novel progressive dual-branch transformer-based diffusion model that explicitly considers pose uncertainty. Our model integrates conventional pose estimation and segmentation methods to initialize distributions, which accelerate the denoising process. A transformer with a cross-attention module is introduced to facilitate information sharing between the pose and mask branches, leading to improved refinement of noisy poses. Experimental results on the MS COCO, 3DMPB, and CrowdPose datasets demonstrate that our approach outperforms baseline methods, achieving a significant increase in average precision (AP). Our dual-branch transformer-based diffusion model not only reduces pose uncertainty but also enhances the efficiency of the reverse diffusion process by utilizing initial noise distributions derived from initial heatmaps and masks. This work highlights the potential of diffusion models in addressing the challenges of 2D human pose estimation under occlusion. The code is available at https://github.com/wenhappyall/ProDiffPose.