Scholarlink

<p>The monocular video-based methods have demonstrated satisfactory performance in recovering plausible 3D human mesh in recent years. However, most of the existing methods cannot effectively utilize non-local information to obtain long-range dependencies, which leads to incorrect alignment of the recovered human mesh. To address this problem, we propose a Region Graph Attention Transformer (RGAT) network for 3D human mesh recovery from monocular video. We introduce a principled region graph representation. It partitions regions related to joint motion by the analysis of predicted feature map information, and attaches attention to semantically meaningful connections. Specifically, we first propose a region-based feature attention module to establish semantic connections among keypoints in a regional manner and recalibrate attention distribution. The introduction of prior knowledge can help RGAT learn the changes of posture. Then, we further develop a dynamic spatial-temporal fusion module to model the continuous representation of motion features. At the same time, we present a hierarchical integration strategy to enhance the influence of adjacent frames on the current frame. Our RGAT achieves superior performance compared with the existing video-based methods on the 3DPW, MPI-INF-3DHP, and Human3.6M datasets.</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

A region Graph attention transformer network for human mesh recovery from monocular video

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Jiaxuan Wang,
  • Shigang Liu,
  • Hanqiang Liu,
  • Yali Peng

摘要

The monocular video-based methods have demonstrated satisfactory performance in recovering plausible 3D human mesh in recent years. However, most of the existing methods cannot effectively utilize non-local information to obtain long-range dependencies, which leads to incorrect alignment of the recovered human mesh. To address this problem, we propose a Region Graph Attention Transformer (RGAT) network for 3D human mesh recovery from monocular video. We introduce a principled region graph representation. It partitions regions related to joint motion by the analysis of predicted feature map information, and attaches attention to semantically meaningful connections. Specifically, we first propose a region-based feature attention module to establish semantic connections among keypoints in a regional manner and recalibrate attention distribution. The introduction of prior knowledge can help RGAT learn the changes of posture. Then, we further develop a dynamic spatial-temporal fusion module to model the continuous representation of motion features. At the same time, we present a hierarchical integration strategy to enhance the influence of adjacent frames on the current frame. Our RGAT achieves superior performance compared with the existing video-based methods on the 3DPW, MPI-INF-3DHP, and Human3.6M datasets.