CLEVRTEX

CLEVRTEX是由牛津大学的视觉几何组创建的一个用于无监督多对象分割的新基准数据集。该数据集包含50000个合成场景，每个场景包含3到10个具有多样形状、纹理和基于物理渲染材料的物体。此外，还有一个包含10000张图像的测试集，使用25种不同的材料。CLEVRTEX旨在通过增加视觉场景的复杂性，推动无监督多对象分割方法的发展，特别是在处理真实世界图像中的纹理和复杂光照效果方面。数据集还包括深度、法线、阴影和因子变化等多种元数据，以及用于从头生成数据集的代码。

CLEVRTEX is a novel benchmark dataset for unsupervised multi-object segmentation, created by the Visual Geometry Group at the University of Oxford. It contains 50,000 synthetic scenes, each comprising 3 to 10 objects with diverse shapes, textures, and physically-based rendered materials. Additionally, there is a test set consisting of 10,000 images that uses 25 distinct materials. CLEVRTEX aims to advance the development of unsupervised multi-object segmentation methods by increasing the complexity of visual scenes, particularly in handling textures and complex lighting effects in real-world images. The dataset also includes various metadata such as depth maps, normal maps, shadows, and factor variations, as well as the code for generating the dataset from scratch.