Dataset Card for widedepth_195fov_300mm_center

This is a FiftyOne dataset with 101 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/widedepth")

# Launch the App
session = fo.launch_app(dataset)

WideDepth in FiftyOne

Educational reference for WideDepth as imported into FiftyOne in this project. The full benchmark is large and multi-configuration; here we document what the dataset contains, which slice we keep, and how it is represented in FiftyOne.

Sources:

• Project page
• Hugging Face dataset
• Video presentation (~3 min overview of the benchmark and pipeline)

---

What is WideDepth?

WideDepth is an indoor fisheye depth-estimation benchmark (ICRA 2026) built from high-resolution LiDAR scans. The authors render synthetic stereo RGB, ground-truth depth, and ground-truth disparity so that pinhole and fisheye models can be compared under controlled geometry.

From the project page:

Property	Detail
Scenes	101 indoor scenes
Labeled pairs	~5K high-resolution stereo pairs with millimeter-accurate depth and disparity
Camera types	Fisheye and pinhole projections of the same content
Variability	Multiple fields of view (FOV), focal lengths, and stereo baselines
Stereo layouts	Horizontal and vertical stereo setups (full benchmark; first vertical fisheye training set at scale)

Why it matters for robotics: Fisheye sensors are common for near-field manipulation and navigation, but indoor benchmarks with trustworthy dense GT were scarce. WideDepth targets metric depth/disparity at wide angles, not just plausible-looking maps.

---

Subset used in this project

We import one camera configuration across all 101 scenes:

Parameter	Value
FOV	195FOV (widest; highlighted in stereo evals on the project page)
Focal length	300mm
Stereo position	CENTER

Files per scene at this path: Up to 8 PNGs on the Hub for this config:

Role	Files
RGB	fisheye.png, pano.png, pano_crop.png
Depth	depth/fisheye.png, depth/pano.png, depth/pano_crop.png
Disparity	disparity/pano.png, disparity/pano_crop.png (no fisheye disparity in this config)

FiftyOne dataset structure

Groups (101)

Each group = one scene at 195FOV / 300mm / CENTER.

Group: scene_id = "001_057_000"
├── slice "fisheye"   → RGB + depth heatmap
├── slice "pano"      → RGB + depth + disparity heatmaps  (default slice)
└── slice "pano_crop" → RGB + depth + disparity heatmaps

FiftyOne concept	Value
group field	fo.Group() with slices fisheye, pano, pano_crop
Default slice	pano
Groups	101
Samples	301 (101 × 3 views − 2 missing pano_crop RGB)

Switching slices in the App shows the same scene under fisheye vs panoramic projections without duplicating scene metadata.

Per-sample fields

Field	Type	Description
filepath	str	Path to RGB PNG (fisheye, pano, or pano_crop)
group	fo.Group element	Slice name within the scene group
scene_id	str	Scene folder name, e.g. 001_057_000
fov	str	Always 195FOV in this import
focal_mm	str	Always 300mm
position	str	Always CENTER
view	str	fisheye | pano | pano_crop (redundant with group slice; useful for filtering)

Labels (heatmaps)

Field	Type	Source file	Present on
depth	fo.Heatmap	depth/{view}.png	All three RGB views
disparity	fo.Heatmap	disparity/{view}.png	pano, pano_crop only

Citation [optional]

When publishing work that uses WideDepth, cite the ICRA 2026 paper from the project page (PDF linked there). BibTeX to be added when available on the page.

---

Related links

• WideDepth project page
• Hugging Face: IlyaInd/WideDepth
• Video presentation
• FiftyOne depth estimation guide