DARTset

手是人类生产力和智慧的载体，最近因3D数字化身的热潮而备受关注。在不同的手部变形模型中，MANO已广泛应用于各种视觉和图形任务。然而，MANO忽略了纹理和配件，这在很大程度上限制了它合成逼真和生活方式手部数据的能力。在本文中，我们将MANO扩展为更加多样化的配件和丰富的纹理，即DART。DART由325张精美的手工制作的纹理图组成，这些纹理图在外观上各不相同，涵盖了不同种类的瑕疵、化妆和配件。我们还提供了Unity GUI，它允许人们用用户特定的设置来渲染手，例如姿势、相机、背景、照明和DART的纹理。通过这种方式，我们生成了大规模(800K)、多样化和高保真的手部图像，并与完美对齐的3D标签(称为DARTset)配对。实验证明了该方法在通用性和多样性方面的优越性。作为现有数据集的一个很好的补充，达特集可以促进手姿态估计和表面重建任务。DART和Unity软件可公开用于研究目的。

Hands are the carrier of human productivity and intelligence, and have recently garnered significant attention amid the boom of 3D digital avatars. Among various hand deformation models, MANO has been widely adopted in diverse vision and graphics tasks. However, MANO neglects textures and accessories, which largely limits its capability to synthesize realistic and lifestyle-aligned hand data. In this paper, we extend MANO to incorporate more diverse accessories and rich textures, namely DART. DART consists of 325 exquisitely handcrafted texture maps that vary in appearance, covering various types of imperfections, makeup, and accessories. We also provide a Unity GUI that enables users to render hands with custom settings including pose, camera, background, lighting, and DART textures. Through this pipeline, we generate a large-scale (800K), diverse, and high-fidelity hand image dataset paired with perfectly aligned 3D labels, referred to as DARTset. Experiments validate the superiority of our approach in terms of generality and diversity. As a valuable supplement to existing datasets, DARTset can facilitate research on hand pose estimation and surface reconstruction tasks. Both DART and the Unity software are publicly available for research purposes.