Scholarlink

The research and application of 2D-to-3D image conversion technology serve as a crucial foundation, significantly enhancing UAV systems’ intelligence, autonomy, safety, and operational efficiency within the aerial access field. In this study, we address the problem of 3D human mesh recovery from single RGB images by building upon the Human Mesh Recovery (HMR) framework. We propose two novel extensions: HMR \(_{KTD}\) , which incorporates hierarchical regression based on kinematic tree structures, and HMR \(_{TFM}\) , which integrates an attention-based mechanism for pose estimation. Although HMR \(_{TFM}\) faces challenges related to overfitting, HMR \(_{KTD}\) demonstrates improved performance over the original HMR. Furthermore, we explore the HMR \(_{INT}\) model, which leverages pretrained Vision Transformers (ViT) and independent prior tokens for pose, shape, and camera parameters. Experimental results on public datasets (3DPW, MPI-INF-3DHP, and Human 3.6M) validate the effectiveness of our approaches.

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

An Efficient Approach for Synthesizing 3D Human Models from 2D Moving Camera Images

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xuan Toan Mai,
  • Thanh Phuong Le,
  • Hong Tai Tran,
  • Tuan-Anh Tran

摘要

The research and application of 2D-to-3D image conversion technology serve as a crucial foundation, significantly enhancing UAV systems’ intelligence, autonomy, safety, and operational efficiency within the aerial access field. In this study, we address the problem of 3D human mesh recovery from single RGB images by building upon the Human Mesh Recovery (HMR) framework. We propose two novel extensions: HMR \(_{KTD}\) , which incorporates hierarchical regression based on kinematic tree structures, and HMR \(_{TFM}\) , which integrates an attention-based mechanism for pose estimation. Although HMR \(_{TFM}\) faces challenges related to overfitting, HMR \(_{KTD}\) demonstrates improved performance over the original HMR. Furthermore, we explore the HMR \(_{INT}\) model, which leverages pretrained Vision Transformers (ViT) and independent prior tokens for pose, shape, and camera parameters. Experimental results on public datasets (3DPW, MPI-INF-3DHP, and Human 3.6M) validate the effectiveness of our approaches.