PASMVS: a dataset for multi-view stereopsis training and reconstruction applications

A large collection of high precision synthetic, path-traced renderings for use in multi-view stereopsis and 3D reconstruction applications. The material properties are primarily highly specular (reflective metals). Ground truth depth maps, model geometry, object masks and camera extrinsic and intrinsic data is provided together with the rendered images. A total of 18,000 samples are provided (45 camera views for 400 scenes), varying the illumination using five, high-definition HDR environment textures, four models (teapot, bunny, armadillo and dragon), ten material properties (bricks, brushed metal, ceramic, checkerboard, concrete, copper, grunge metal, marble, piano/ivory, and steel) and two camera focal lengths (35 mm and 50 mm). File descriptions: + PASMVS.blend - Blender file (developed using Blender version 2.8.1) used to generate the dataset. The Blender file packs only one of the HDR environmental textures to use as an example. + PASMVS_sample.jpg - a visual collage of 8 scenes, illustrating the variability introduced by using different models, illumination, material properties and camera focal lengths. + PASMVS_576p.tar.xz - The full PASMVS dataset containing all 400 scenes (768x576 pixel resolution). Note that tar.xz archive used due to file size limitations (7.8 Gb); uncompressed the total file size is approximately 35 Gb. The "index.csv" in the root folder provides a summary for all 18,000 scenes, specifying the corresponding file names alongside the the subfolder. + PFMdemo.zip - Python example source code for loading ground truth depth map (PFM file format) as a numpy array. + modelSTL - Stereolithography (STL) file for the models used in the dataset. + sample_armadillo1bricks50mm.zip - A single PASMVS scene sample for easy viewing and download.本数据集为高精度合成路径追踪渲染图像的大型集合，适用于多视图立体匹配（multi-view stereopsis, MVS）与三维重建任务。其场景材质属性以高镜面反射特性（反光金属类）为主。随渲染图像一同提供的配套数据包括真值深度图、模型几何结构、目标掩码，以及相机外参与内参数据。 本次数据集共包含18000个样本（400个场景，每个场景对应45个相机视角），通过以下变量组合生成：5款高清高动态范围（High Dynamic Range, HDR）环境纹理用于改变照明条件，4种基础模型（茶壶、兔子、犰狳与龙），10种材质属性（砖块、拉丝金属、陶瓷、棋盘格、混凝土、铜、做旧金属、大理石、钢琴/象牙及钢材），以及2种相机焦距（35 mm与50 mm）。 文件说明如下： + PASMVS.blend：用于生成该数据集的Blender工程文件（基于Blender 2.8.1版本开发），该文件仅打包了一款HDR环境纹理作为示例。 + PASMVS_sample.jpg：包含8个场景的可视化拼贴图，展示了不同模型、照明条件、材质属性与相机焦距带来的效果差异。 + PASMVS_576p.tar.xz：完整的PASMVS数据集，包含全部400个场景，图像分辨率为768×576像素。考虑文件大小限制，本数据集采用tar.xz归档格式（归档大小7.8 GB），解压后总容量约为35 GB。根目录下的index.csv文件汇总了全部18000个样本，列出了各样本对应的文件名与所属子文件夹。 + PFMdemo.zip：用于加载真值深度图（PFM文件格式）并转换为NumPy数组的Python示例源码。 + modelSTL：数据集中所用模型的立体光刻（STL）文件。 + sample_armadillo1bricks50mm.zip：单组PASMVS场景样本，便于用户快速查看与下载。