structured3d-spatiallm-pzhang

## Note by pzhang

1. Door and window heights are recorded as their actual dimensions, rather than the full wall height provided in the original data.
2. Room heights are determined independently, instead of using the single global height defined for the entire scene in the original data.

# Structured3D-SpatialLM Dataset

Structured3D dataset preprocessed in SpatialLM format for layout estimation with LLMs.

## Overview

This dataset is derived from [Structured3D](https://structured3d-dataset.org/) **3,500 synthetic house designs** created by professional designers, preprocessed and formatted specifically for [SpatialLM](https://github.com/manycore-research/SpatialLM) training.

Point clouds and layouts are derived from the [RoomFormer](https://github.com/ywyue/RoomFormer/tree/main/data_preprocess) data preprocessing script.

## Data Extraction

Point clouds and layouts are compressed in zip files. To extract the files, run the following script:

```bash
cd structured3d-spatiallm
chmod +x extract.sh
./extract.sh
```

## Dataset Strucutre

```bash
structured3d-spatiallm/
├── structured3d_train.json # Training conversations
├── structured3d_test.json # Test conversations
├── dataset_info.json # Dataset metadata
├── split.csv # Train/val split mapping
├── train.csv # Train info
├── test.csv # val info
├── pcd/ # Point cloud data
│ └── .ply
├── layout/ # Scene layout annotations
│ └── .txt
└── extract.sh # Extraction script
```

The `structured3d_train.json` and `structured3d_test.json` dataset follows the **SpatialLM format** with ShareGPT-style conversations:

```json
{
"conversations": [
{
"from": "human",
"value": "<point_cloud>Detect walls, doors, windows. The reference code is as followed: ..."
},
{
"from": "gpt",
"value": "<|layout_s|>wall_0=...<|layout_e|>"
}
],
"point_clouds": ["pcd/scene_id.ply"]
}
```

## License

This dataset is derived from [Structured3D](https://github.com/bertjiazheng/Structured3D) dataset. Please refer to the original dataset's license terms for usage restrictions.

## Citation

If you use this dataset in your research, please cite the original Structured3D paper:

```bibtex
@inproceedings{Structured3D,
title = {Structured3D: A Large Photo-realistic Dataset for Structured 3D Modeling},
author = {Jia Zheng and Junfei Zhang and Jing Li and Rui Tang and Shenghua Gao and Zihan Zhou},
booktitle = {Proceedings of The European Conference on Computer Vision (ECCV)},
year = {2020}
}
```

pzhang 注：
1. 门窗高度记录为其实际尺寸，而非原始数据中提供的完整墙体高度。
2. 房间高度将独立确定，而非使用原始数据中为整个场景定义的单一全局高度。

# Structured3D-SpatialLM 数据集

已按照SpatialLM格式预处理的Structured3D数据集，用于大语言模型（Large Language Model）的布局估计任务。

## 概述
本数据集源自[Structured3D](https://structured3d-dataset.org/) **3500套由专业设计师创作的合成住宅设计方案**，经过专门预处理并格式化，以适配[SpatialLM](https://github.com/manycore-research/SpatialLM)的训练需求。
点云数据与布局信息源自[RoomFormer](https://github.com/ywyue/RoomFormer/tree/main/data_preprocess)的数据预处理脚本。

## 数据解压
点云与布局文件均以ZIP压缩包形式存储。如需解压文件，请运行以下脚本：
bash
cd structured3d-spatiallm
chmod +x extract.sh
./extract.sh

## 数据集结构
bash
structured3d-spatiallm/
├── structured3d_train.json # 训练对话数据
├── structured3d_test.json # 测试对话数据
├── dataset_info.json # 数据集元数据
├── split.csv # 训练/验证集划分映射表
├── train.csv # 训练集信息
├── test.csv # 验证集信息
├── pcd/ # 点云数据目录
│ └── .ply 格式文件
├── layout/ # 场景布局标注目录
│ └── .txt 格式文件
└── extract.sh # 解压脚本

`structured3d_train.json` 与 `structured3d_test.json` 遵循**SpatialLM格式**，采用ShareGPT风格的对话格式：
json
{
"conversations": [
{
"from": "human",
"value": "<point_cloud>Detect walls, doors, windows. The reference code is as followed: ..."
},
{
"from": "gpt",
"value": "<|layout_s|>wall_0=...<|layout_e|>"
}
],
"point_clouds": ["pcd/scene_id.ply"]
}

## 许可证
本数据集源自[Structured3D](https://github.com/bertjiazheng/Structured3D)数据集，请遵循原始数据集的许可证条款使用本数据集。

## 引用
若您在研究中使用本数据集，请引用原始Structured3D论文：
bibtex
@inproceedings{Structured3D,
title = {Structured3D: A Large Photo-realistic Dataset for Structured 3D Modeling},
author = {Jia Zheng and Junfei Zhang and Jing Li and Rui Tang and Shenghua Gao and Zihan Zhou},
booktitle = {Proceedings of The European Conference on Computer Vision (ECCV)},
year = {2020}
}