FreiHAND
收藏资源简介:
从单个 RGB 图像估计 3D 手部姿势是一个高度模糊的问题,它依赖于无偏的训练数据集。在本文中,我们分析了在现有数据集上进行训练时的跨数据集泛化。我们发现这些方法在他们训练的数据集上表现良好,但不能推广到其他数据集或野外场景。因此,我们引入了第一个伴随 3D 手部姿势和形状注释的大规模多视图手部数据集。为了注释这个真实世界的数据集,我们提出了一种迭代的、半自动的“人在环”方法,其中包括手部拟合优化,以推断每个样本的 3D 姿势和形状。我们表明,在我们的数据集上训练的方法在其他数据集上测试时始终表现良好。此外,该数据集允许我们训练一个网络,该网络可以从单个 RGB 图像中预测完整的关节手形。
Estimating 3D hand pose from a single RGB image is a highly ill-posed problem relying on unbiased training datasets. In this paper, we analyze cross-dataset generalization when training on existing datasets. We find that these methods perform well on the datasets they are trained on, but fail to generalize to other datasets or in-the-wild scenarios. Therefore, we introduce the first large-scale multi-view hand dataset accompanied by 3D hand pose and shape annotations. To annotate this real-world dataset, we propose an iterative, semi-automatic "human-in-the-loop" method that includes hand fitting optimization to infer the 3D pose and shape for each sample. We demonstrate that methods trained on our dataset perform consistently well when tested on other datasets. Furthermore, this dataset enables us to train a network that can predict full articulated hand shapes from a single RGB image.

- FreiHAND数据集首次发表,由Christian Zimmermann等人提出,旨在提供高质量的手部姿态估计数据。
- FreiHAND数据集首次应用于手部姿态估计任务,展示了其在计算机视觉领域的潜力。
- FreiHAND数据集被广泛用于手部姿态估计和手势识别的研究,成为该领域的重要基准数据集之一。



