RenderIH

RenderIH是一个大规模的合成数据集，专门用于3D交互手势姿态估计。该数据集由阿里巴巴集团和上海人工智能实验室创建，包含100万张照片级真实图像，涵盖了多样的背景、视角和手部纹理。数据集通过新的姿态优化算法生成自然且多样的交互姿态，并引入基于Transformer的姿态估计网络TransHand，以提高姿态估计的准确性。RenderIH数据集不仅适用于提高任何手势姿态估计方法的准确性，还能显著降低误差，从6.76mm降至5.79mm。该数据集的应用领域包括人机交互、增强现实、虚拟现实和手语识别等，旨在解决3D交互手势姿态从单一RGB图像中获取的挑战。

RenderIH is a large-scale synthetic dataset specifically designed for 3D interactive hand gesture pose estimation. Created by Alibaba Group and the Shanghai AI Laboratory, it contains 1 million photorealistic images covering diverse backgrounds, viewpoints and hand textures. The dataset generates natural and diverse interactive poses via a novel pose optimization algorithm, and introduces the Transformer-based pose estimation network TransHand to improve the accuracy of pose estimation. The RenderIH dataset can not only be applied to enhance the accuracy of any hand gesture pose estimation method, but also significantly reduce the error from 6.76 mm to 5.79 mm. Its application fields include human-computer interaction, augmented reality, virtual reality and sign language recognition, and it aims to address the challenge of recovering 3D interactive hand gesture poses from a single RGB image.