README.md

---
license: mit
task_categories:
- image-to-3d
configs:
- config_name: doors
  data_files: doors/*.json
- config_name: stairs
  data_files: stairs/*/*.ply
- config_name: windows
  data_files: windows/*.json
- config_name: poses
  data_files: poses/*.json
- config_name: structures
  data_files: structures/*.obj
---

# HouseLayout3D: A Benchmark Dataset for 3D Layout Estimation in the Wild

[Paper](https://huggingface.co/papers/2512.02450) | [Project Page](https://houselayout3d.github.io) | [Code](https://github.com/HouseLayout3D/houselayout3d)

**HouseLayout3D** is a challenging benchmark dataset for **3D layout estimation in large-scale, multi-floor buildings**. It is built upon real-world building scans from [Matterport3D](https://niessner.github.io/Matterport/), and provides detailed annotations of structural elements across up to five floors and forty rooms per building. The dataset is designed to support research in scene understanding, indoor mapping, and robotics applications that require vectorized, object-free representations of indoor spaces.

<img src="visualization.jpg" alt="Visualization of HouseLayout3D Scenes" width="400"/>

## Why HouseLayout3D?

Most existing datasets and models for 3D layout estimation are tailored for small-scale, single-room scenes. These models, while effective on synthetic or room-scale data, fail to generalize to larger, multi-level buildings. HouseLayout3D addresses this gap by providing:

- **Real-world, complex architectural spaces**, including open-plan layouts, staircases, and multi-floor transitions with up to **5 floors** and **40+ rooms**.
- **Detailed structural annotations**, including walls, floors, ceilings, doors, windows, and stairs.

## What’s in the Dataset?

- Vectorized layout annotations of structures (ceilings, walls, floors) as 3D polygons.
- Annotations of windows and doors (including opening direction) as 3D rectangles.
- Annotations of stairs as 3D polygons.

## Sample Usage

Follow these steps to set up the environment, download the dataset, and visualize a scene.

### 1. Installation

First, clone the main repository and install dependencies:

```bash
conda create --name houselayout3d python=3.10 -y
conda activate houselayout3d

git clone https://github.com/HouseLayout3D/houselayout3d.git
cd houselayout3d
pip install -r requirements.txt

# Install Git LFS to download large files
# On Mac
brew install git-lfs
# On Linux
sudo apt-get install git-lfs
```

### 2. Download the Dataset

Download the HouseLayout3D dataset using `git lfs` from Hugging Face:

```bash
git lfs install # Ensure lfs is initialized
git clone https://huggingface.co/datasets/houselayout3d/HouseLayout3D data
```

### 3. Visualize a Scene

Navigate to the `houselayout3d` directory (where you cloned the code in step 1) and run the visualization script:

```bash
python visualize.py
python -m http.server 6008
```
Then open your browser and navigate to `http://localhost:6008` to view the visualizations.

## Data Structure
```text
HouseLayout3D/
├── structures/
│   ├── {scene}.obj
│   │   → Base layout as a triangle mesh (without doors or windows)
│   └── layouts_split_by_entity/
│       └── {scene}/*.obj
│           → Individual wall / ceiling / floor entities (jointly compose the base layout)
├── doors/
│   └── {scene}.json
│       → Door annotations: each door is a list of 4 corners. The "normal" field indicates opening direction.
├── windows/
│   └── {scene}.json
│       → Window annotations: each window is a list of 4 corners.
├── stairs/
│   └── {scene}/*.ply
│       → Stair annotations as polygonal surfaces that follow the steps (steps themselves are part of the base layout)
├── poses/
│   └── {scene}.json
│       → The camera poses & intrinsics used for evaluation of the layout depth in nerfstudio format
```