---
license: cc-by-4.0
task_categories:
- depth-estimation
language:
- en
tags:
- endoscopy
- colonoscopy
- depth
- pose
- medical
dataset_info:
  features:
  - name: dataset_name
    dtype: string
  - name: sequence
    dtype: string
  - name: frame_idx
    dtype: int32
  - name: frame_idx_prev
    dtype: int32
  - name: frame_idx_curr
    dtype: int32
  - name: frame_idx_next
    dtype: int32
  - name: rgb_prev
    dtype: image
  - name: rgb_curr
    dtype: image
  - name: rgb_next
    dtype: image
  - name: depth
    dtype: image
  - name: occlusion
    dtype: image
  - name: pose_curr2prev
    sequence: float64
  - name: pose_curr2next
    sequence: float64
  - name: K
    sequence: float32
  - name: has_depth
    dtype: bool
  - name: has_occlusion
    dtype: bool
  - name: has_pose
    dtype: bool
  - name: source_fps
    dtype: float32
  - name: target_fps
    dtype: float32
  - name: frame_stride
    dtype: int32
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train/**/*.parquet
  - split: val
    path: data/val/**/*.parquet
  - split: test
    path: data/test/**/*.parquet
---

# C3VDv2 — Colonoscopy 3D Video Dataset v2

This dataset is a re-packaged version of **C3VDv2** originally published by
Johns Hopkins University, distributed under the
[Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/) license.

**Original dataset DOI:** https://doi.org/10.7281/T1/JC64MK
**Dataset archive:** https://archive.data.jhu.edu/dataset.xhtml?persistentId=doi:10.7281/T1/JC64MK

## Attribution

This re-packaged version was created to facilitate streaming access. The original
data and all intellectual property rights belong to the original authors and
Johns Hopkins University. When using this dataset, you must comply with the
CC BY 4.0 license terms, which require attribution to the original creators.

Please cite the original work (check the dataset page above for the full citation):

```bibtex
@dataset{c3vdv2,
  title  = {C3VDv2: Colonoscopy 3D Video Dataset v2},
  doi    = {10.7281/T1/JC64MK},
  url    = {https://archive.data.jhu.edu/dataset.xhtml?persistentId=doi:10.7281/T1/JC64MK},
  publisher = {Johns Hopkins University Data Archive},
}
```

## Re-packaging

Raw omnidirectional fisheye frames (1350×1080) are converted to undistorted
perspective crops (512×512) using the Scaramuzza camera model from
`camera_intrinsics.txt`. Each **Parquet row** is one training sample:

| column | dtype | description |
|--------|-------|-------------|
| `sequence` | string | sequence name, e.g. `c1_ascending_t1_v1` |
| `frame_idx` | int32 | centre frame index |
| `frame_idx_prev/curr/next` | int32 | original frame indices in the triplet |
| `rgb_prev` | image | undistorted 512×512 PNG (HF Image feature) |
| `rgb_curr` | image | undistorted 512×512 PNG (HF Image feature) |
| `rgb_next` | image | undistorted 512×512 PNG (HF Image feature) |
| `depth` | image | 16-bit PNG, 512×512, uint16 value → metres: `val / 65535 * 0.1` |
| `occlusion` | image | 8-bit PNG, 512×512, 255 = occluded, 0 = clear |
| `pose_curr2prev` | list[float64] | 16-value row-major 4×4 relative pose |
| `pose_curr2next` | list[float64] | 16-value row-major 4×4 relative pose |
| `K` | list[float32] | 9-value row-major 3×3 normalised camera intrinsics |
| `has_depth/has_occlusion/has_pose` | bool | supervised label availability flags |
| `frame_stride` | int32 | original-frame stride between triplet neighbours |

## Splits

| split | trials | description |
|-------|--------|-------------|
| train | t1, t2 | two trajectories per region |
| val   | t4     | held-out trajectory |
| test  | t3     | held-out trajectory |

## Usage

```python
from datasets import load_dataset
import numpy as np
from PIL import Image
import io

ds = load_dataset("SmartWhatt/c3vdv2-SfM", split="train", streaming=True)

for row in ds:
    rgb   = np.array(row["rgb_curr"])                                    # 512×512×3 uint8
    depth = np.array(row["depth"]).astype(np.float32) / 65535.0 * 0.1  # 512×512 metres
    occ   = np.array(row["occlusion"]) > 0                              # 512×512 bool mask
    T_curr2prev = np.array(row["pose_curr2prev"]).reshape(4, 4)
    K = np.array(row["K"]).reshape(3, 3)
```