ShapeNet

ShapeNet is an ongoing effort to establish a richly-annotated, large-scale dataset of 3D shapes. We provide researchers around the world with this data to enable research in computer graphics, computer vision, robotics, and other related disciplines. ShapeNet is a collaborative effort between researchers at Princeton, Stanford and TTIC. ShapeNetCore is a densely annotated subset of ShapeNet covering 55 common object categories with ~51,300 unique 3D models. Each model in ShapeNetCore are linked to an appropriate synset in WordNet (version 3.0). The distribution of for ShapeNetCore is organized into one zip file per synset. Each zip file is named by the synset noun offset as an eight-digit zero padded string. For example, bench is contained within 02828884.zip since the WordNet synset offset for bench is 02828884 (you can browse WordNet 3.1 online). The corresponding ImageNet synsets can be accessed at http://www.image-net.org/synset?wnid=n where is replaced by the padded synset offset (note that ImageNet includes an 'n' prefix for noun synsets). For instance, the ImageNet url for bench is http://www.image-net.org/synset?wnid=n02828884. ShapeNetCore v2 is an update to ShapeNetCore with improved quality of model geometry and fixed issues relating to materials and textures. The v2 package provides improved .obj+.mtl format model data that replaces the .obj+.mtl data in ShapeNetCore v1, and is part of an upcoming public data release. As part of ShapeNetCore v2, surface and solid voxelizations (using binvox) of the models are also provided. However, the preliminary ShapeNetCore v2 was incomplete and had different normalization than ShapeNetCore v1. NOTE: The OBJ files are now pre-aligned so that the up direction is the +Y axis, and the front is the -Z axis (this fixes the mirroring issue in ShapeNetCore v1). The Y-Z plane is the bilateral symmetry plane for most categories. NOTE: The OBJ files have been pre-aligned so that the up direction is the +Y axis, and the front is the +X axis. In addition each model is diagonally normalized to fit within a unit cube centered at the origin. The X-Y plane is the bilateral symmetry plane for most categories. A known issue with ShapeNetCore v1 is that due to the conversion pipeline used, some models are mirrored and have incorrect normals. ShapeNetSem is a subset of ShapeNet, richly annotated with physical attributes, which we release for the benefit of the research community. Compared to ShapeNetCore, ShapeNetSem is a smaller, more densely annotated subset of ShapeNet consisting of 12,000 models spread over a broader set of 270 categories. There are several types of metadata available for the models. For more information, see the ShapeNetSem v0 README.

ShapeNet是一项持续推进的研究工作，旨在构建一个标注丰富的大规模三维形状数据集。我们向全球研究者开放该数据集，以推动计算机图形学、计算机视觉、机器人学及其他相关学科的研究。ShapeNet由普林斯顿大学、斯坦福大学与丰田技术学院（TTIC）的研究者合作完成。

ShapeNetCore是ShapeNet的密集标注子集，涵盖55个常见物体类别，包含约51300个独特三维模型。ShapeNetCore中的每个模型均与WordNet（3.0版）中的同义词集（synset）相关联。ShapeNetCore的数据集按每个同义词集打包为单独的zip压缩包，每个压缩包以同义词集名词偏移量（八位零填充字符串）命名。例如，“长椅”对应的WordNet同义词集偏移量为02828884，因此其数据包为02828884.zip（可在线浏览WordNet 3.1）。对应的ImageNet同义词集可通过网址http://www.image-net.org/synset?wnid=n[偏移量]访问（注：ImageNet的名词同义词集前缀为“n”）。例如，长椅对应的ImageNet网址为http://www.image-net.org/synset?wnid=n02828884。

ShapeNetCore v2是ShapeNetCore的更新版本，优化了模型几何质量，修复了材质与纹理相关的问题。v2数据包提供了改进后的.obj+.mtl格式模型数据，替代了ShapeNetCore v1中的.obj+.mtl数据，属于即将公开的数据集发布内容之一。此外，ShapeNetCore v2还提供了模型的表面体素化与实体体素化结果（采用binvox工具生成）。不过早期的ShapeNetCore v2版本并不完整，且与ShapeNetCore v1的归一化方式存在差异。

注意：当前的OBJ文件已完成预对齐，朝上方向为+Y轴，朝前方向为-Z轴（此修正解决了ShapeNetCore v1中的镜像问题），多数类别的双边对称平面为Y-Z平面。另一说明：OBJ文件已完成预对齐，朝上方向为+Y轴，朝前方向为+X轴。此外，每个模型均经过对角线归一化，以适配以原点为中心的单位立方体，多数类别的双边对称平面为X-Y平面。

ShapeNetCore v1存在一处已知问题：由于所采用的转换流水线限制，部分模型存在镜像问题且法向量有误。

ShapeNetSem是ShapeNet的子集，附带丰富的物理属性标注，我们向研究社区开放该数据集以助力相关研究。相较于ShapeNetCore，ShapeNetSem规模更小，但标注更为密集，包含12000个模型，覆盖270个更广泛的类别。该数据集为模型提供了多种类型的元数据。更多信息请参阅ShapeNetSem v0自述文件。

搜集汇总

背景与挑战

背景概述

ShapeNet是一个大规模、丰富标注的3D形状数据集，主要用于计算机图形学、计算机视觉和机器人等领域的研究。它包含ShapeNetCore和ShapeNetSem两个子集，分别涵盖55个常见物体类别和270个更广泛的类别，并具有不同版本和特点。