ShapeNetSDF
收藏资源简介:
ShapeNetSDF是一个从ShapeNet Core衍生而来的有符号距离场点样本数据集,专为训练和评估3D形状上的隐式神经表示(神经场)而设计。该数据集将每个3D形状转换为水密流形,归一化到单位立方体[-1, 1]³内,并采样生成三种点集:uniform(在[-1,1]³内均匀采样的262144个点及其精确SDF值,用于监督自由空间中的场)、surface(在网格表面附近添加高斯噪声采样的262144个点,其SDF值接近但不完全为零,用于表面附近的训练)和groundtruth(精确的表面点,SDF值约为零,用于评估,如表面重建)。每个点样本存储为一个形状为[262144, 4]的float32数组,包含x, y, z坐标和有签距离值(SDF)。数据集包含51个常见的3D对象类别(如椅子、桌子、飞机等),总大小约为512 GB。数据按类别组织,并提供了训练/验证/测试分割(约80%/10%/10%)以及用于快速实验的精选子集。数据集遵循ShapeNet的使用条款,仅限非商业研究使用。
ShapeNetSDF is a signed distance field point sample dataset derived from ShapeNet Core, specifically designed for training and evaluating implicit neural representations (neural fields) on 3D shapes. The dataset converts each 3D shape into a watertight manifold, normalizes it within the unit cube [-1, 1]³, and samples three types of point sets: uniform (262144 points uniformly sampled within [-1,1]³ with exact SDF values, for supervising the field in free space), surface (262144 points sampled with Gaussian noise near the mesh surface, with SDF values close but not exactly zero, for training near the surface), and groundtruth (exact surface points with SDF values approximately zero, for evaluation such as surface reconstruction). Each point sample is stored as a float32 array with shape [262144, 4], containing x, y, z coordinates and the signed distance value (SDF). The dataset includes 51 common 3D object categories (e.g., chairs, tables, airplanes, etc.), with a total size of approximately 512 GB. Data is organized by category and provides train/validation/test splits (approximately 80%/10%/10%) as well as curated subsets for quick experimentation. The dataset follows ShapeNets terms of use and is limited to non-commercial research purposes.
数据集概述
ShapeNetSDF 是一个从 ShapeNet Core 导出的带符号距离场(SDF)点样本数据集,主要用于训练和评估三维形状上的隐式神经表示(Neural Fields / Implicit Neural Representations)。
数据规模与结构
- 数据总量:约 512 GB。
- 类别数量:51 个类别,包括
airplane,chair,table,car,bed等常见物体,以及一个汇总所有类别的all/文件夹。 - 单个模型:每个 ShapeNet 模型被转换为水密流形,并归一化到单位立方体
[-1, 1]³内。 - 点集类型:每个模型包含三个点集,每个点集存储为一个
[N, 4]的float32数组,四列分别代表x,y,z坐标和 SDF 值(负内正外):- uniform:均匀采样点,共 262,144 个(64³),用于监督自由空间的场。
- surface:添加高斯噪声的近表面采样点,SDF 值接近零,用于表面附近训练。
- groundtruth:精确的表面点,SDF ≈ 0,用于评估(如表面重建、Chamfer 距离、IoU)。
数据集目录结构
ShapeNetSDF/ ├── <category>/ # 51 个类别文件夹 │ ├── uniform/<model_id>.npy │ ├── surface/<model_id>.npy │ ├── groundtruth/<model_id>.npy │ ├── train.txt # ~80% 划分 │ ├── val.txt # ~10% 划分 │ └── test.txt # ~10% 划分 ├── all/ # 跨所有类别的汇总 │ ├── uniform/ surface/ groundtruth/ │ ├── train.txt / val.txt / test.txt # 全局划分(种子 42) │ ├── 10k10c.txt / 5k10c.txt / 5k5c.txt / 100_5c.txt # 精选子集(模型数量_类别数c) │ └── labels.json # 类别 ↔ 模型ID 映射 └── README.md
数据划分
- train.txt / val.txt / test.txt:每个类别文件夹及
all/文件夹均包含,比例为 ~80% / 10% / 10%(种子 42)。 - 精选子集(位于
all/下):10k10c.txt,5k10c.txt,5k5c.txt,100_5c.txt:用于快速实验的少量模型和类别组合。
创建方法
- 从 ShapeNet Core 获取原始模型(
models/model_normalized.obj)。 - 使用
point_cloud_utils将网格转换为水密流形。 - 归一化到以原点为中心的单位立方体内(缩放比例为最大顶点距离的 0.98)。
- 采样
uniform、surface(含噪声)和groundtruth点集。 - 计算每个点的带符号距离,保存为
[N, 4]的float32数组。 - 处理过程基于
model_id设定随机种子,确保可复现。
许可证
该数据集衍生自 ShapeNet,使用需遵守 ShapeNet 使用条款,仅限非商业研究用途。




