Scholarlink

<p>In computer vision, real-time and photorealistic 3D hand rendering plays a critical role in human-computer interaction. However, existing Neural Radiance Field (NeRF)-based methods often struggle with the intricate geometry and self-occlusion of human hands, limiting both realism and efficiency. In this paper, we present HoloHand, a hybrid framework that enhances real-time neural hand rendering by integrating self-occlusion-aware appearance modeling and depth-aware learning. Built upon an explicit hand mesh representation, our method captures high-frequency appearance details via a neural implicit field while maintaining structural accuracy. To capture self-occlusion-induced appearance variations under dynamic hand poses—such as inter-finger shadowing—we introduce a self-occlusion-aware appearance field that learns to predict per-point irradiance attenuation from training data. Additionally, a depth-guided supervision loss is incorporated to improve geometry consistency, and a super-resolution module enables high-resolution output while maintaining inference speed within the real-time regime. Extensive experiments demonstrate that HoloHand achieves high-quality rendering results in real time, bridging the gap between performance and photorealism in neural hand rendering.</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

HoloHand: enhancing real-time neural hand rendering with self-occlusion-aware appearance fields and depth-aware learning

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Qinghan Xiao,
  • Kai Lu,
  • Menglei Zhang,
  • Qiang Li,
  • Wankou Yang

摘要

In computer vision, real-time and photorealistic 3D hand rendering plays a critical role in human-computer interaction. However, existing Neural Radiance Field (NeRF)-based methods often struggle with the intricate geometry and self-occlusion of human hands, limiting both realism and efficiency. In this paper, we present HoloHand, a hybrid framework that enhances real-time neural hand rendering by integrating self-occlusion-aware appearance modeling and depth-aware learning. Built upon an explicit hand mesh representation, our method captures high-frequency appearance details via a neural implicit field while maintaining structural accuracy. To capture self-occlusion-induced appearance variations under dynamic hand poses—such as inter-finger shadowing—we introduce a self-occlusion-aware appearance field that learns to predict per-point irradiance attenuation from training data. Additionally, a depth-guided supervision loss is incorporated to improve geometry consistency, and a super-resolution module enables high-resolution output while maintaining inference speed within the real-time regime. Extensive experiments demonstrate that HoloHand achieves high-quality rendering results in real time, bridging the gap between performance and photorealism in neural hand rendering.