FreiHAND
收藏资源简介:
FreiHAND数据集是由德国弗莱堡大学和Adobe研究机构合作创建的大型多视角手部数据集,包含33000条记录,旨在解决单张RGB图像中手部姿态和形状的3D估计问题。该数据集通过创新的迭代半自动化‘人机交互’方法进行标注,包括手部拟合优化以推断每个样本的3D姿态和形状。数据集涵盖32个不同的人,具有完全可动的手部形状,高变化的手部姿态,并包括与物体的互动。部分数据集,标记为训练集,是在绿色屏幕前捕捉的,便于与不同背景图像合成。测试集包含在不同室内外环境中的记录。FreiHAND数据集不仅提高了跨数据集泛化能力,还允许训练网络从单张RGB图像预测完整的手部3D形状,为手部形状估计提供了一个挑战性的基准。
The FreiHAND Dataset is a large-scale multi-view hand dataset co-created by the University of Freiburg in Germany and Adobe Research, containing 33,000 records. It aims to solve the 3D estimation problem of hand pose and shape from a single RGB image. This dataset is annotated via an innovative iterative semi-automatic "human-computer interaction" method, which includes hand fitting optimization to infer the 3D pose and shape of each sample. The dataset covers 32 distinct individuals, features fully articulated hand shapes, highly varied hand poses, and includes interactions with objects. A portion of the dataset, designated as the training set, was captured in front of a green screen to facilitate compositing with different background images. The test set contains recordings taken in various indoor and outdoor environments. The FreiHAND Dataset not only improves cross-dataset generalization but also enables training neural networks to predict complete 3D hand shapes from single RGB images, providing a challenging benchmark for hand shape estimation.

- 1FreiHAND: A Dataset for Markerless Capture of Hand Pose and Shape from Single RGB Images弗莱堡大学 · 2019年



