SpatialGen-Testset

# SpatialGen Testset

This repository contains the test set for [SPATIALGEN: Layout-guided 3D Indoor Scene Generation](https://arxiv.org/abs/2509.14981), a novel multi-view multi-modal diffusion model for generating realistic and semantically consistent 3D indoor scenes.

[Project page](https://manycore-research.github.io/SpatialGen) | [Paper](https://arxiv.org/abs/2509.14981) | [Code](https://github.com/manycore-research/SpatialGen)

We provide a test set of 48 preprocessed point clouds and their corresponding GT layouts, multi-view images are cropped from the high-resolution panoramic images.

## Folder Structure

Outlines of the dataset files:
```bash
SpatialGen-Testset
|-- {scene_id} # scene folder
| |-- rgb # multi-view color images
| |-- semantic # multi-view semantic maps
| |-- depth # multi-view depth maps
| |-- normal # multi-view normal maps
| |-- layout_depth # rendered depth maps of the GT layout
| |-- layout_semantic # rendered semantic maps of the GT layout
| |-- cameras.json # camera parameters for the scene
| |-- room_layout.json # GT structured layout in JSON format
│ └-- test_split.txt # testing split
└── test_split_caption.jsonl # textural captions for each scene
```

## Sample Usage

### Visualization

We provide a [code](https://github.com/manycore-research/SpatialGen/blob/main/visualize_layout.py) to visualize the layout data.

```python
data_root_dir = "./example_data/scenes"
vis_root_dir = "./visualization"
scene_data_dirs = [os.path.join(data_root_dir, d) for d in os.listdir(data_root_dir) if d.startswith("scene_")]

for scene_data_dir in scene_data_dirs:
vis_output_dir = os.path.join(vis_root_dir, scene_data_dir.split('/')[-1])
# save layout_bbox.ply and camera poses in vis_output_dir
visualize_spatialgen_data(scene_data_dir, vis_output_dir)
```

### Inference

This dataset is used for evaluating the SpatialGen models for 3D indoor scene generation. The following commands from the [code repository](https://github.com/manycore-research/SpatialGen) demonstrate how to run inference for different tasks (after following the installation instructions in the repository).

**Single image-to-3D Scene Generation**

```bash
bash scripts/infer_spatialgen_i2s.sh
```

**Text-to-image-to-3D Scene Generation**

You can choose a pair of `scene_id` and `prompt` from `captions/spatialgen_testset_captions.jsonl` to run the text-to-scene experiment.

```bash
bash scripts/infer_spatialgen_t2s.sh
```

# SpatialGen测试集
本仓库包含针对[SPATIALGEN：布局引导的三维室内场景生成模型（SPATIALGEN: Layout-guided 3D Indoor Scene Generation）](https://arxiv.org/abs/2509.14981)的测试集，该模型是一款用于生成真实且语义一致的三维室内场景的新型多视图多模态扩散模型。

[项目主页](https://manycore-research.github.io/SpatialGen) | [论文](https://arxiv.org/abs/2509.14981) | [代码](https://github.com/manycore-research/SpatialGen)

本测试集包含48组经过预处理的点云及其对应的真值（ground truth）布局，多视图图像均从高分辨率全景图像中裁剪得到。

## 文件夹结构
数据集文件结构如下：
bash
SpatialGen-Testset
|-- {scene_id} # 场景文件夹
| |-- rgb # 多视图彩色图像
| |-- semantic # 多视图语义分割图
| |-- depth # 多视图深度图
| |-- normal # 多视图法向图
| |-- layout_depth # 真值布局的渲染深度图
| |-- layout_semantic # 真值布局的渲染语义分割图
| |-- cameras.json # 场景相机参数文件
| |-- room_layout.json # JSON格式的真值结构化布局文件
│ └-- test_split.txt # 测试划分文件
└── test_split_caption.jsonl # 各场景文本描述文件

## 示例用法
### 可视化
我们提供了一段[代码](https://github.com/manycore-research/SpatialGen/blob/main/visualize_layout.py)用于可视化布局数据。
python
data_root_dir = "./example_data/scenes"
vis_root_dir = "./visualization"
scene_data_dirs = [os.path.join(data_root_dir, d) for d in os.listdir(data_root_dir) if d.startswith("scene_")]

for scene_data_dir in scene_data_dirs:
vis_output_dir = os.path.join(vis_root_dir, scene_data_dir.split('/')[-1])
# save layout_bbox.ply and camera poses in vis_output_dir
visualize_spatialgen_data(scene_data_dir, vis_output_dir)

### 推理
本数据集用于评估用于三维室内场景生成的SpatialGen模型。下述代码仓库中的命令展示了如何针对不同任务运行推理（需先按照仓库中的安装说明完成环境配置）。

**单图像转三维场景生成**
bash
bash scripts/infer_spatialgen_i2s.sh

**文本-图像-三维场景生成**
您可以从`captions/spatialgen_testset_captions.jsonl`中选取一组`scene_id`与`prompt`来运行文本转场景实验。
bash
bash scripts/infer_spatialgen_t2s.sh