README.md

---
language:
- en
pretty_name: WideDepth train
task_categories:
- depth-estimation
tags:
- autonomous-driving
- fisheye
- equirectangular
- lidar
- disparity
- depth-estimation
size_categories:
- 10K<n<100K
license: cc-by-nc-4.0
---

# WideDepth train

## Dataset Summary
WideDepth train is an outdoor multi-view fisheye training dataset for depth/disparity learning and domain adaptation.

The dataset accompanies the paper *WideDepth: Millimeter-Accurate Benchmark for Fisheye Depth Estimation*, accepted to ICRA 2026.

It provides synchronized frames across:
- left fisheye RGB
- right fisheye RGB
- upper equirect RGB
- lower virtual equirect RGB
- equirect disparity
- fisheye depth

This subset was captured outdoors with a handheld rig (fisheye stereo + LiDAR), then rectified/warped to equirect views and paired with sparse LiDAR-derived labels.

## Dataset Structure
Expected layout:

```text
WideDepth_train/
  left_fisheye/
  right_fisheye/
  equirect_up/
  equirect_down_virt/
  disp_equirect/
  depth_fisheye/
  manifests/
```

All modality folders use synchronized zero-padded numeric names (for example `00000.png`, `00001.png`, ...).

Content by folder:
- `left_fisheye`, `right_fisheye`, `equirect_up`, `equirect_down_virt`: RGB images (`uint8`, 3-channel)
- `disp_equirect`: disparity map image files (typically scaled by `100`)
- `depth_fisheye`: projected sparse depth map image files in left fisheye frame (typically depth in `mm`)
- `manifests`: filtering/reindexing outputs and mapping files

## Source Data and Processing
High-level pipeline:
1. Outdoor handheld capture with a synchronized fisheye stereo + LiDAR rig.
2. Stereo rectification and warping to equirectangular views.
3. Projection of merged LiDAR points to the camera frame to generate sparse depth and derived disparity targets.
4. Release preparation with optional RGB anonymization (face/plate blur) and synchronized frame indexing across modalities.

Projection model:
- Equirectangular warping and LiDAR-to-image projection are performed with the **Double Sphere** camera model using calibrated intrinsics/extrinsics.

Capture setup (train release):
- Two ZED X One fisheye cameras in a vertical stereo configuration.
- Camera lens FoV is approximately 180° horizontal and 120° vertical; calibration was done at `1920x1080`.
- Two Livox Mid-360 LiDARs with partially overlapping FoVs.
- Typical sensor rates: cameras at 30 Hz and LiDARs at 10 Hz, aligned by timestamp in ROS.
- Data was captured outdoors (city areas, parks, streets), daylight, mostly cloudy conditions.

Notes:
- Reported train-set sparse projected depth density is approximately 8.1%.
- This release is intended as training data; create your own deterministic train/val/test split if needed.

## Limitations and Privacy
- `depth_fisheye` is sparse and can be noisy, especially on thin/far structures.
- Depth overlays can show local drift due to sparsity, timing offsets, and calibration residuals.
- Disparity/depth are image-encoded and require proper scaling in loaders.

Anonymization in this release is **best-effort**:
- All published RGB images in this release were processed with automated face/plate blurring.
- Automated blur reduces identifiable content but may miss cases.
- Verify privacy compliance for your jurisdiction and use case before redistribution/deployment.

## Links
- Paper: https://arxiv.org/abs/2605.24074
- WideDepth benchmark dataset: https://huggingface.co/datasets/IlyaInd/WideDepth
- WideDepth code repository: https://github.com/IlyaInd/WideDepth

## License
This release uses **Creative Commons Attribution-NonCommercial 4.0 International** (`cc-by-nc-4.0`).

You must provide attribution and may not use the dataset for commercial purposes.

## Citation
```bibtex
@article{indyk2026widedepth,
  title={WideDepth: Millimeter-Accurate Benchmark for Fisheye Depth Estimation},
  author={Indyk, Ilia and Penshin, Ignat and Sosin, Ivan and Monastyrny, Maxim and Valenkov, Aleksei and Makarov, Ilya},
  journal={arXiv preprint arXiv:2605.24074},
  year={2026}
}
```