cs-mshah/SynMirror

SynMirror是一个大规模的合成数据集，专注于镜面反射，包含多种镜子类型、物体、相机姿态、HDRI背景和地板纹理。数据集由两个广泛使用的3D对象数据集（Objaverse和Amazon Berkeley Objects）的3D资产渲染而成，总共有198,204个渲染样本。每个渲染样本包含颜色、类别ID分割图、深度、法线和相机状态等信息。数据集由Vision and AI Lab, IISc Bangalore策划，采用Open Data Commons License Attribution family许可。数据集的未来版本将随机化物体位置和旋转，并修复某些家具对象的接地问题。

SynMirror is a first-of-its-kind large-scale synthetic dataset focused on mirror reflections. It is created by rendering 3D assets from two widely used 3D object datasets, Objaverse and Amazon Berkeley Objects (ABO), in a virtual blender environment. The dataset contains 198,204 rendered samples, each with attributes like colors, category_id_segmaps, depth, normals, and cam_states. The dataset is curated by the Vision and AI Lab at IISc Bangalore and is licensed under the Open Data Commons License Attribution family. Future versions of the dataset aim to randomize object locations and rotations, and fix minor issues.