---
pretty_name: EndoX Dataset
tags:
- endoscopy
- synthetic dataset
- physics-based rendering
- webdataset
---

# EndoX Dataset

## Folder Structure

Each sequence is stored as one `.tar` WebDataset shard.

```text
sample_dataset/
  colon/
    sequence_1/
      sample_dataset-colon-sequence_1.tar
  small_intestines/
    sequence_1/
      sample_dataset-small_intestines-sequence_1.tar
    sequence_2/
      sample_dataset-small_intestines-sequence_2.tar
dataset/
  colon/
    sequence_1/
      dataset-colon-sequence_1.tar
    sequence_2/
      dataset-colon-sequence_2.tar
    sequence_3/
      dataset-colon-sequence_3.tar
    sequence_4/
      dataset-colon-sequence_4.tar
    sequence_5/
      dataset-colon-sequence_5.tar
```

Inside each shard, every frame is one WebDataset sample:

```text
0000.depth.png
0000.normal_camera.png
0000.normal_world.png
0000.occlusion.png
0000.optical_flow.png
0000.rgb.png
```

Sequence metadata is embedded in each shard:

```text
__metadata__/camera_pose/camera_pose.txt
__metadata__/coverage_map/coverage_vertices.npz
__metadata__/coverage_map/coverage_vertices_per_frame.npz
```

## How To Load The Dataset

Install dependencies:

```bash
pip install webdataset pillow
```

Stream one sequence directly from Hugging Face:

```python
import webdataset as wds

url = (
    "https://huggingface.co/datasets/bealam111/EndoX_Dataset/resolve/"
    "main/dataset/colon/sequence_1/dataset-colon-sequence_1.tar"
)

dataset = (
    wds.WebDataset(url)
    .select(lambda sample: "rgb.png" in sample)
    .decode("pil")
)

for sample in dataset:
    frame_id = sample["__key__"]
    rgb = sample["rgb.png"]
    depth = sample["depth.png"]
    normal_camera = sample["normal_camera.png"]
    normal_world = sample["normal_world.png"]
    occlusion = sample["occlusion.png"]
    optical_flow = sample["optical_flow.png"]
    break
```

Download one sequence with the Hugging Face CLI:

```bash
hf download bealam111/EndoX_Dataset \
  --type dataset \
  --include "dataset/colon/sequence_1/dataset-colon-sequence_1.tar" \
  --local-dir ./EndoX_Dataset
```

## Optional: Extract To Modality Folders

The `.tar` files use WebDataset names such as `0000.rgb.png` so each frame can be streamed as one multi-modal sample. After downloading, users who prefer a traditional folder structure can reorganize one extracted shard into modality folders:

```python
from pathlib import Path
import shutil

sequence_dir = Path("./EndoX_Dataset/dataset/colon/sequence_1")
tar_path = sequence_dir / "dataset-colon-sequence_1.tar"
extract_dir = sequence_dir / "extracted"

shutil.unpack_archive(tar_path, extract_dir)

for path in list(extract_dir.glob("*.png")):
    frame_id, modality, ext = path.name.split(".", 2)
    target = extract_dir / modality / f"{frame_id}.{ext}"
    target.parent.mkdir(parents=True, exist_ok=True)
    path.rename(target)
```

After running the script, the extracted sequence has this structure:

```text
extracted/
  rgb/
    0000.png
    0001.png
  depth/
    0000.png
    0001.png
  normal_camera/
    0000.png
  normal_world/
    0000.png
  occlusion/
    0000.png
  optical_flow/
    0000.png
  __metadata__/
    camera_pose/camera_pose.txt
    coverage_map/coverage_vertices.npz
    coverage_map/coverage_vertices_per_frame.npz
```