Scholarlink

<p>Reconstructing two interacting hands from a monocular RGB image presents a formidable challenge due to similar appearances and mutual occlusions. To address these issues, we propose the Extract-Regress Network (<b>ER</b>), a unified framework comprising specialized modules for visual feature extraction, feature fusion, and 3D mesh vertex regression. The Extract module includes DeSim for decoupling and capturing appearance details of separate hands and DeOcc for processing latent connections and spatial clues from interacting hands. The Regress module employs FuseJoint to enhance feature representation by fusing joint position messages into visual feature maps. Our approach achieves state-of-the-art performance on the InterHand2.6M dataset, with a mean per joint position error (MPJPE) of 6.65 mm, outperforming existing methods by significant margins. This work advances the field of image-based 3D hand reconstruction, offering robust solutions for virtual reality, augmented reality, and human–computer interaction applications. Our paper code is available at <a href="https://github.com/Cantherine101424/ER">https://github.com/Cantherine101424/ER</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

ER: Extract-regress network for precise 3D reconstruction of interacting hands from monocular images

  • 期刊论文

  • 发表日期：

    2026年1月4日

查看全文
  • Rui Feng,
  • Fangni Chen,
  • Dongsheng Xu,
  • Linzhan Zhong

摘要

Reconstructing two interacting hands from a monocular RGB image presents a formidable challenge due to similar appearances and mutual occlusions. To address these issues, we propose the Extract-Regress Network (ER), a unified framework comprising specialized modules for visual feature extraction, feature fusion, and 3D mesh vertex regression. The Extract module includes DeSim for decoupling and capturing appearance details of separate hands and DeOcc for processing latent connections and spatial clues from interacting hands. The Regress module employs FuseJoint to enhance feature representation by fusing joint position messages into visual feature maps. Our approach achieves state-of-the-art performance on the InterHand2.6M dataset, with a mean per joint position error (MPJPE) of 6.65 mm, outperforming existing methods by significant margins. This work advances the field of image-based 3D hand reconstruction, offering robust solutions for virtual reality, augmented reality, and human–computer interaction applications. Our paper code is available at https://github.com/Cantherine101424/ER.