README.md

---
language:
- en
license: cc-by-nc-4.0
size_categories:
- 100M<n<1B
pretty_name: InfraDepth
task_categories:
- depth-estimation
tags:
- 3d-point-cloud
- image-restoration
- image-segmentation
- civil-engineering
---

## InfraDepth

`InfraDepth` is a multimodal dataset of rendered depth map patches for masonry bridges and tunnels.  
It is designed to support research on **image restoration, inpainting, sparse-to-dense depth reconstruction, and segmentation** of civil infrastructure components.  

The dataset combines **3D point clouds of masonry bridges and tunnels**, projected through a virtual camera into patches, then stored as `.npz` files with depth maps, masks, and camera parameters.  

---

**Paper**: [InfraDiffusion: zero-shot depth map restoration with diffusion models and prompted segmentation from sparse infrastructure point clouds](https://huggingface.co/papers/2509.03324)

**Code**: [https://github.com/Jingyixiong/InfraDiffusion-official-implement](https://github.com/Jingyixiong/InfraDiffusion-official-implement)

---

## 📁 Dataset Structure

```bash
datasets/
│
├── masonry_bridges/
│   ├── begc/
│   │   ├── arch/
│   │   │   ├── 0/
│   │   │   │   └── rendered_0.8_0.8_0.5/
│   │   │   │       ├── patch_0.npz
│   │   │   │       ├── patch_0_cam_params.npz
│   │   │   │       └── ...
│   │   ├── pier/
│   │   └── spandrel_wall/
│   │
│   └── hertfordshire/
│       ├── arch/
│       ├── pier/
│       └── spandrel_wall/
│
└── tunnels/
    └── wheatly_tunnel/
        ├── S-15/
        │   ├── arch/
        │   └── pier/
        ├── S-20/
        │   ├── arch/
        │   └── pier/
        └── S-25/
            ├── arch/
            └── pier/
```

Each component folder (for example, `arch/0/`) contains a folder named `rendered_0.8_0.8_0.5/` where the patches are stored.  
The suffix `0.8_0.8_0.5` indicates the patch bounding box size in meters (x, y, z).  

---

## 🔹 File Formats

Inside each `rendered_0.8_0.8_0.5/` folder:

```bash
| File name                  | Format | Description |
|-----------------------------|--------|-------------|
| patch_{idx}.npz            | NPZ    | Contains depth map and masks |
| patch_{idx}_cam_params.npz | NPZ    | Camera intrinsics and extrinsics for the patch |

Each `patch_{idx}.npz` file contains:

- `depth_map`: Rendered depth values(original depth map without image restoration)
- `mask_inpainting`: Mask region for inpainting  
- `mask_boundary`: Boundary mask of the patch  
```

---

## ✨ Sample Usage

The `InfraDepth` dataset is designed to be used with the `InfraDiffusion` framework. Below are examples from the official GitHub repository on how to run InfraDiffusion restoration using the dataset:

**(1) Masonry Tunnel Dataset**
```bash
python main.py data=tunnels \
    image_restore.deg=inpainting \
    image_restore.sigma_y=0.16 \
    general.save_results=true
```

**(2) Masonry Bridge Dataset**
```bash
python main.py data=masonry_bridges \
    image_restore.deg=inpainting \
    image_restore.sigma_y=0.16 \
    general.save_results=true
```

**(3) Selecting a Specific Infrastructure (infrastructure names can be found in `configs/data`)**
Example: To just get image restoration results on `hertfordshire`, override it:
```bash
python main.py \
    data=masonry_bridges \
    data.infra_name='begc' \
    image_restore.deg=inpainting \
    image_restore.sigma_y=0.16 \
    general.save_results=true
```

For more detailed usage instructions, including environment setup and SAM segmentation, please refer to the [official GitHub repository](https://github.com/Jingyixiong/InfraDiffusion-official-implement).

---

## 📚 Citation
If you use this dataset, please cite the associated paper:

```bibtex
@article{jing2025infradiffusion,
  title={InfraDiffusion: zero-shot depth map restoration with diffusion models and prompted segmentation from sparse infrastructure point clouds},
  author={Jing, Yixiong and Zhang, Cheng and Wu, Haibing and Wang, Guangming and Wysocki, Olaf and Sheil, Brian},
  year={2025},
  note={Preprint}
}
```