Scholarlink

<p>By modeling the contact relationships between human body parts and local object regions from images, potential cues underlying human–object interactions can be revealed, thereby alleviating the uncertainty in affordance learning caused by the ambiguity of possible actions. However, existing methods often follow a “contact region segmentation–transfer” paradigm that relies solely on appearance-level mappings, neglecting the semantic and geometric consistency of human–object interactions. As a result, when facing significant variations in objects or diverse collaborative relationships among body parts, the model’s perception and prediction capabilities are degraded. To this end, we propose a visual-geometric collaborative guided affordance learning network that jointly leverages visual and geometric cues to model the complex relationships in human–object interactions and the contact feature representation. Particularly, a semantic-pose heuristic perception module is devised to exploit both semantic and geometric cues to guide the network to focus on interaction-relevant regions, improving contact-region perception under various interaction scenarios. Meanwhile, a geometric-apparent alignment transfer module is introduced to jointly align local regions of apparent and structural similarity, eliminating the transport difficulties posed by intra-class correspondence ambiguity. Furthermore, we construct a larger and more diverse Contact &amp; Object-Part Affordance dataset, consisting of 55,047 images covering 61 object categories and 35 affordance types, enabling a more comprehensive evaluation of the model’s perception and prediction capabilities across complex interaction scenarios. Experimental results demonstrate that our method outperforms the representative models regarding objective metrics and visual quality. Project: <a href="https://github.com/lhc1224/VCR-Net">github.com/lhc1224/VCR-Net</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

Visual-Geometric Collaborative Guidance for Affordance Learning

  • 期刊论文

  • 发表日期：

    2026年7月22日

查看全文
  • Hongchen Luo,
  • Wei Zhai,
  • Jiao Wang,
  • Yang Cao,
  • Zheng-Jun Zha

摘要

By modeling the contact relationships between human body parts and local object regions from images, potential cues underlying human–object interactions can be revealed, thereby alleviating the uncertainty in affordance learning caused by the ambiguity of possible actions. However, existing methods often follow a “contact region segmentation–transfer” paradigm that relies solely on appearance-level mappings, neglecting the semantic and geometric consistency of human–object interactions. As a result, when facing significant variations in objects or diverse collaborative relationships among body parts, the model’s perception and prediction capabilities are degraded. To this end, we propose a visual-geometric collaborative guided affordance learning network that jointly leverages visual and geometric cues to model the complex relationships in human–object interactions and the contact feature representation. Particularly, a semantic-pose heuristic perception module is devised to exploit both semantic and geometric cues to guide the network to focus on interaction-relevant regions, improving contact-region perception under various interaction scenarios. Meanwhile, a geometric-apparent alignment transfer module is introduced to jointly align local regions of apparent and structural similarity, eliminating the transport difficulties posed by intra-class correspondence ambiguity. Furthermore, we construct a larger and more diverse Contact & Object-Part Affordance dataset, consisting of 55,047 images covering 61 object categories and 35 affordance types, enabling a more comprehensive evaluation of the model’s perception and prediction capabilities across complex interaction scenarios. Experimental results demonstrate that our method outperforms the representative models regarding objective metrics and visual quality. Project: github.com/lhc1224/VCR-Net.