PlankAssembly

# PlankAssembly Dataset

If you encounter downloading issue, you can directly download the dataset [here](https://manycore-research-azure.kujiale.com/manycore-research/PlankAssembly/data.zip).

## Dataset Description

- **Homepage:** https://manycore-research.github.io/PlankAssembly
- **Repository:** https://github.com/manycore-research/PlankAssembly
- **Paper:** https://arxiv.org/abs/2308.05744

### Dataset Summary

This is the dataset used for training [PlankAssembly](https://manycore-research.github.io/PlankAssembly). It contains 26,707 shape programs derived from parametric CAD models.

## Dataset Structure

PlankAssembly dataset is a directory with the following structure:

PlankAssemblyDataset
├── model # shape program
| └── <MODLE_ID>.json
└── splits # dataset splits
├── train.txt
├── valid.txt
└── test.txt

## PlankAssembly DSL

A cabinet is typically assembled by a list of plank models, where each plank is represented as an axis-aligned cuboid. A cuboid has six degrees of freedom, which correspond to the starting and ending coordinates along the three axes:
```
Cuboid (x_min, y_min, z_min, x_max, y_max, z_max).
```
Each coordinate can either take a numerical value or be a pointer to the corresponding coordinate of another cuboid (to which it attaches to).

In the parametric modeling software, a plank is typically created by first drawing a 2D profile and then applying the extrusion command. Thus, we categorize the faces of each plank into *sideface* or *endface*, depending on whether they are along the direction of the extrusion or not. Then, given a pair of faces from two different planks, we consider that an attachment relationship exists if (i) the two faces are within a distance threshold of 1mm and (ii) the pair consists of one sideface and one endface.

## Shape Program

Each shape program (*model.json*) is a JSON file with the following structure:

```python
{
# model id
"name": str,
# numerical values of all planks, the units are millimeters
"planks": List[List], # N x 6
# extrusion direction of each plank
"normal": List[List], # N x 3
# attachment relationships
# -1 denotes no attachment relationship
# Others denote the index of the flattened plank sequence
"attach": List[List], # N x 6
}
```

## BibTex

Please cite our paper if you use PlankAssembly dataset in your work:

```bibtex
@inproceedings{PlankAssembly,
author = {Hu, Wentao and Zheng, Jia and Zhang, Zixin and Yuan, Xiaojun and Yin, Jian and Zhou, Zihan},
title = {PlankAssembly: Robust 3D Reconstruction from Three Orthographic Views with Learnt Shape Programs},
booktitle = {ICCV},
year = {2023}
}
```

# 木板装配（PlankAssembly）数据集

若遇到下载问题，可直接通过以下链接下载数据集：https://manycore-research-azure.kujiale.com/manycore-research/PlankAssembly/data.zip。

## 数据集说明

- **官方主页：** https://manycore-research.github.io/PlankAssembly
- **代码仓库：** https://github.com/manycore-research/PlankAssembly
- **相关论文：** https://arxiv.org/abs/2308.05744

### 数据集概述

本数据集用于训练[木板装配（PlankAssembly）](https://manycore-research.github.io/PlankAssembly)模型，包含26707个源自参数化计算机辅助设计（Computer Aided Design, CAD）模型的形状程序。

## 数据集结构

木板装配数据集采用目录结构，具体组织形式如下：

PlankAssemblyDataset
├── model # 形状程序
| └── <模型ID>.json
└── splits # 数据集划分
├── train.txt
├── valid.txt
└── test.txt

## 木板装配领域专用语言（PlankAssembly DSL）

柜体通常由一系列木板模型装配而成，每块木板可表示为轴对齐长方体。长方体具有6个自由度，对应沿三个坐标轴的起始与终止坐标：

长方体（x_min, y_min, z_min, x_max, y_max, z_max）。

每个坐标既可以取数值，也可以指向另一个长方体的对应坐标（二者以此实现连接）。

在参数化建模软件中，木板通常通过先绘制二维轮廓再执行拉伸命令来创建。因此，我们根据面是否沿拉伸方向，将每块木板的面分为*侧面（sideface）*与*端面（endface）*两类。对于来自两块不同木板的一对面，若满足以下两个条件，则认为二者存在装配连接关系：(1) 两个面的距离小于1毫米的阈值；(2) 该面组由一个侧面和一个端面组成。

## 形状程序

每个形状程序（即`model.json`文件）为JSON格式文件，结构如下：

python
{
# 模型ID
"name": str,
# 所有木板的数值参数，单位为毫米
"planks": List[List], # N × 6
# 每块木板的拉伸方向
"normal": List[List], # N × 3
# 装配连接关系
# -1 表示无装配连接关系
# 其他值表示扁平化木板序列的索引
"attach": List[List], # N × 6
}

## BibTex引用

若在研究工作中使用木板装配数据集，请引用以下论文：

bibtex
@inproceedings{PlankAssembly,
author = {Hu, Wentao and Zheng, Jia and Zhang, Zixin and Yuan, Xiaojun and Yin, Jian and Zhou, Zihan},
title = {PlankAssembly: Robust 3D Reconstruction from Three Orthographic Views with Learnt Shape Programs},
booktitle = {ICCV},
year = {2023}
}

搜集汇总

背景与挑战

背景概述

PlankAssembly是一个用于3D重建的数据集，包含26,707个从参数化CAD模型衍生的形状程序，主要用于训练PlankAssembly模型。数据集以JSON格式组织，详细描述了木板的几何参数、法线方向和附件关系，支持从三个正交视图进行稳健的3D重建。