S3O4D (Stanford 3D Objects for Disentangling)

实验结果的“Stanford 3D Objects”部分中使用的数据可在 TensorFlow Datasets 中获得。数据由来自斯坦福 3D 扫描存储库的兔子和龙对象的 100,000 个渲染组成。未来可能会添加更多对象，但论文中只使用了兔子和龙。每个对象都使用来自 2 球体上的一个点的均匀采样照明和均匀采样的 3D 旋转来渲染。真正的潜在状态与图像一起作为 NumPy 数组提供。光照以具有单位范数的 3 向量给出，而旋转以四元数和 3x3 正交矩阵的形式提供。

The data utilized in the "Stanford 3D Objects" section of the experimental results is accessible via TensorFlow Datasets. This dataset consists of 100,000 renderings of the bunny and dragon objects sourced from the Stanford 3D Scanning Repository. While additional objects may be added in future iterations, only the bunny and dragon were employed in this paper. Each object is rendered using uniformly sampled illumination from a point on two spheres and uniformly sampled 3D rotations. The ground-truth latent states are provided alongside the corresponding images as NumPy arrays. The illumination is represented as a 3-dimensional unit norm vector, while the rotations are supplied in two forms: quaternions and 3×3 orthogonal matrices.

搜集汇总

背景与挑战

背景概述

S3O4D是由DeepMind于2020年发布的数据集，包含来自斯坦福3D扫描存储库的兔子和龙对象的100,000个渲染图像，每个图像均附带均匀采样的照明和3D旋转信息。潜在状态以NumPy数组形式提供，包括单位范数的光照向量以及四元数和正交矩阵表示的旋转数据。