MesaTask-10K

<h2 align="center"<strong>MesaTask: Towards Task-Driven Tabletop Scene Generation via 3D Spatial Reasoning</strong></h2>

<div align="center">
<a href='https://huggingface.co/datasets/InternRobotics/MesaTask-10K'><img src='https://img.shields.io/badge/Paper-arXiv-%232986fc'></a> &nbsp;
<a href='https://github.com/InternRobotics/MesaTask'><img src='https://img.shields.io/badge/Code-GitHub-%23181717?&logo=github'></a> &nbsp;
<a href='https://mesatask.github.io/'><img src='https://img.shields.io/badge/Home-Website-05a4a7?'></a> &nbsp;
</div>

<p align="center">
<img src="assets/teaser_final.png" alt="MesaTask Teaser" width="800"/>
</p>

## 🔑Key Features
MesaTask-10K, a large-scale dataset for task-oriented tabletop scene generation, comprises approximately 10,700 synthetic tabletop scenes across 6 common indoor table types, along with an asset library of over 12,000 3D objects (covering more than 200 classes) each with detailed semantic information.

## 🔥Get Started
#### Download the Dataset

To download the full dataset, you can use the following code. If you encounter any issues, please refer to the official Hugging Face documentation.
```shell
# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install

# When prompted for a password, use an access token with write permissions.
# Generate one from your settings: https://huggingface.co/settings/tokens
git clone https://huggingface.co/datasets/InternRobotics/InternRobotics/MesaTask-10K

# If you want to clone without large files - just their pointers
GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/InternRobotics/InternRobotics/MesaTask-10K
```

## Dataset Structure

#### Folder hierarchy

```shell
MesaTask-10K/
|-- Asset_annotation.json
|-- sbert_text_features.pkl
|-- Assets_library/
|-- {uid}.glb
|-- ...
|-- Layout_info/
|-- bathroom_vanity/
|-- bathroom_vanity_0000/
|-- front.png
|-- layout.json
|-- bathroom_vanity_0001/
|-- ...
|-- coffee_table/
|-- dining_table/
|-- dressing_table/
|-- kitchen_counter/
|-- office_table/
```

#### File format

The asset_annotation format is listed as follows:
```shell
{
"523b89a8ef4947a588c7844034aa9dc2": {
"uid": "523b89a8ef4947a588c7844034aa9dc2",
"source": "holodeck",
"category": "coin",
"onFloor": true,
"onObject": true,
"onWall": false,
"onCeiling": false,
"onTable": true,
"detailed_caption": "Circular metallic coin with intricate designs, stylized logo, and shiny reflective surface.",
"mass": "0.005",
"materials": [
"metal",
"copper",
"nickel"
],
"is_container": false,
"textured": true,
"bbox": {
"min": {
"x": -0.9980270266532898,
"y": -0.08568830043077469,
"z": -1.0
},
"max": {
"x": 0.9980270266532898,
"y": 0.08568830043077469,
"z": 1.0
}
}
}
...
}
```

The layout format is listed as follows:
```shell
{
"scene_settings": {
"units": "centimeters",
"up_axis": "Z"
},
"item_placement_zone": [
0.0,
116.8,
0.0,
73.8
],
"restricted_zone": [
33.7,
83.0,
7.8,
40.6
],
"objects": [
{
"name": "7_towels_1_dcda83fe-3261-4606-a29c-c4e6d24701fb",
"instance": "7_towels_1",
"retrieved_uid": "dcda83fe-3261-4606-a29c-c4e6d24701fb",
"original_size": [
0.1894739270210266,
0.5998818278312683,
0.20736036449670792
],
"scale_factor": [
54.7988419074881,
37.26036757747303,
54.79378534573218
],
"rotation": [
0.0,
0.0,
0.0,
1.0
],
"z_rotation": 0.0,
"size": [
10.4,
22.4,
11.3
],
"position": [
91.4,
50.0,
5.5
]
}
...
]
}
```

## ✏️ Citation

```bibtex
@misc{hao2025mesatask,
title={MesaTask: Towards Task-Driven Tabletop Scene Generation via 3D Spatial Reasoning},
author={Jinkun Hao and Naifu Liang and Zhen Luo and Xudong Xu and Weipeng Zhong and Ran Yi and Yichen Jin and Zhaoyang Lyu and Feng Zheng and Lizhuang Ma and Jiangmiao Pang},
year={2025},
eprint={2509.22281},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2509.22281},
}
```

<h2 align="center"><strong>MesaTask：面向三维空间推理的任务驱动桌面场景生成</strong></h2>

<div align="center">
<a href='https://huggingface.co/datasets/InternRobotics/MesaTask-10K'><img src='https://img.shields.io/badge/%E8%AE%BA%E6%96%87-arXiv-%232986fc'></a> &nbsp;
<a href='https://github.com/InternRobotics/MesaTask'><img src='https://img.shields.io/badge/%E4%BB%A3%E7%A0%81-GitHub-%23181717?&logo=github'></a> &nbsp;
<a href='https://mesatask.github.io/'><img src='https://img.shields.io/badge/%E4%B8%BB%E9%A1%B5-%E5%AE%98%E7%BD%91-05a4a7?'></a> &nbsp;
</div>

<p align="center">
<img src="assets/teaser_final.png" alt="MesaTask 预览图" width="800"/>
</p>

## 🔑核心特性
MesaTask-10K是一款面向任务导向桌面场景生成的大规模数据集，包含覆盖6种常见室内桌型的约10700个合成桌面场景，同时附带拥有超过12000个三维物体的资产库（涵盖200余个类别），每个物体均配有详细语义信息。

## 🔥快速上手
### 📥 下载数据集
如需下载完整数据集，可执行以下代码。若遇到任何问题，请参考Hugging Face官方文档。
shell
# 确保已安装git-lfs（https://git-lfs.com）
git lfs install

# 若提示输入密码，请使用带有写入权限的访问令牌。可在以下地址生成令牌：https://huggingface.co/settings/tokens
git clone https://huggingface.co/datasets/InternRobotics/InternRobotics/MesaTask-10K

# 若仅需克隆文件指针而非完整大文件
GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/InternRobotics/InternRobotics/MesaTask-10K

## 📁 数据集结构
### 文件夹层级结构
shell
MesaTask-10K/
|-- Asset_annotation.json
|-- sbert_text_features.pkl
|-- Assets_library/
|-- {uid}.glb
|-- ...
|-- Layout_info/
|-- bathroom_vanity/
|-- bathroom_vanity_0000/
|-- front.png
|-- layout.json
|-- bathroom_vanity_0001/
|-- ...
|-- coffee_table/
|-- dining_table/
|-- dressing_table/
|-- kitchen_counter/
|-- office_table/

### 文件格式
资产标注文件格式如下：
shell
{
"523b89a8ef4947a588c7844034aa9dc2": {
"uid": "523b89a8ef4947a588c7844034aa9dc2",
"source": "holodeck",
"category": "硬币",
"onFloor": true,
"onObject": true,
"onWall": false,
"onCeiling": false,
"onTable": true,
"detailed_caption": "带有精致纹路、定制化标识与光亮反光表面的圆形金属硬币。",
"mass": "0.005",
"materials": [
"金属",
"铜",
"镍"
],
"is_container": false,
"textured": true,
"bbox": {
"min": {
"x": -0.9980270266532898,
"y": -0.08568830043077469,
"z": -1.0
},
"max": {
"x": 0.9980270266532898,
"y": 0.08568830043077469,
"z": 1.0
}
}
}
...
}

场景布局文件格式如下：
shell
{
"scene_settings": {
"units": "厘米",
"up_axis": "Z轴"
},
"item_placement_zone": [
0.0,
116.8,
0.0,
73.8
],
"restricted_zone": [
33.7,
83.0,
7.8,
40.6
],
"objects": [
{
"name": "7_towels_1_dcda83fe-3261-4606-a29c-c4e6d24701fb",
"instance": "7_towels_1",
"retrieved_uid": "dcda83fe-3261-4606-a29c-c4e6d24701fb",
"original_size": [
0.1894739270210266,
0.5998818278312683,
0.20736036449670792
],
"scale_factor": [
54.7988419074881,
37.26036757747303,
54.79378534573218
],
"rotation": [
0.0,
0.0,
0.0,
1.0
],
"z_rotation": 0.0,
"size": [
10.4,
22.4,
11.3
],
"position": [
91.4,
50.0,
5.5
]
}
...
]
}

## ✏️ 引用格式
bibtex
@misc{hao2025mesatask,
title={MesaTask: Towards Task-Driven Tabletop Scene Generation via 3D Spatial Reasoning},
author={Jinkun Hao and Naifu Liang and Zhen Luo and Xudong Xu and Weipeng Zhong and Ran Yi and Yichen Jin and Zhaoyang Lyu and Feng Zheng and Lizhuang Ma and Jiangmiao Pang},
year={2025},
eprint={2509.22281},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2509.22281},
}