Voxel51/widedepth
Dataset Card for WideDepth in FiftyOne FiftyOne dataset for WideDepth — an indoor fisheye depth-estimation benchmark (ICRA 2026) with millimeter-accurate ground-truth depth and disparity rendered from high-resolution LiDAR scans. We use one fixed camera configuration from the full WideDepth benchmark — 195° FOV, 300 mm focal length, CENTER stereo position — across all 101 indoor scenes. The full Hub release has many combinations (4 FOVs × 5 focal lengths × 3 positions, plus… See the full description on the dataset page: https://huggingface.co/datasets/Voxel51/widedepth.
Dataset Card for WideDepth in FiftyOne
FiftyOne dataset for WideDepth — an indoor fisheye depth-estimation benchmark (ICRA 2026) with millimeter-accurate ground-truth depth and disparity rendered from high-resolution LiDAR scans.
We use one fixed camera configuration from the full WideDepth benchmark — 195° FOV, 300 mm focal length, CENTER stereo position — across all 101 indoor scenes.
The full Hub release has many combinations (4 FOVs × 5 focal lengths × 3 positions, plus multiple RGB/depth/disparity views per config). We standardize on the widest FOV setting highlighted in the paper’s hardest evaluations, so every scene shares the same virtual rig instead of mixing 60+ configs per scene.
Installation
If you haven't already, install FiftyOne:
pip install -U fiftyoneUsage
import fiftyone as fo
from huggingface_hub import snapshot_download
# Download the dataset snapshot to the current working directory
snapshot_download(
repo_id="Voxel51/widedepth",
local_dir=".",
repo_type="dataset"
)
# Load dataset from current directory using FiftyOne's native format
dataset = fo.Dataset.from_dir(
dataset_dir=".", # Current directory contains the dataset files
dataset_type=fo.types.FiftyOneDataset, # Specify FiftyOne dataset format
name="WideDepth" # Assign a name to the dataset for identification
)
# Launch the App
session = fo.launch_app(dataset)
Dataset Soureces
- Project page: https://ilyaind.github.io/WideDepth/
- Hugging Face: https://huggingface.co/datasets/IlyaInd/WideDepth
- Paper: https://arxiv.org/pdf/2605.24074v1
- Video overview: https://www.youtube.com/watch?v=zmxO8p5tIB8
About WideDepth
WideDepth provides synthetic stereo RGB, dense depth, and disparity for 101 indoor scenes across multiple camera configurations. This FiftyOne dataset uses a single fixed rig per scene:
Known gaps: Scenes 096_092_000 and 097_092_270 are missing the pano_crop view and have no 3D slice.
Dataset summary
Group structure
Each group is one scene (e.g. 001_057_000) with up to four slices:
WideDepth ships RGB, depth, and disparity only — not ready-made point clouds. For this dataset, we backprojected the pano_crop ground-truth depth (metric, millimeter-accurate) into 3D using the paper’s equirectangular camera model, colored each point from the matching RGB pixel, and packaged the result as an fo3d scene for the FiftyOne 3D viewer. Each cloud is a single-view snapshot from one capture position, not a full room reconstruction.
Group: scene_id = "001_057_000"
├── fisheye → image
├── pano → image
├── pano_crop → image (default)
└── pointcloud → 3dSwitch slices in the App to compare projections or open the 3D view for the same capture.
Sample fields
Labels
Depth and disparity are stored as `fo.Heatmap` labels on the image slices (not as separate samples).
Depth and disparity maps are 16-bit PNG ground truth in millimeters. Zero values indicate invalid or masked regions (common on wide panoramic views).
Citation
@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},
url = {https://arxiv.org/abs/2605.24074}
}