README.md

metadata

license: cc-by-4.0
task_categories:
  - depth-estimation
tags:
  - panorama
  - depth
  - lidar
  - outdoor
  - equirectangular
  - '360'
size_categories:
  - n<1K
pretty_name: ZüriPano
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

🗃️ ZüriPano Dataset

ZüriPano is a real-world outdoor panoramic depth benchmark, captured with the Leica RTC360 LiDAR scanner (8K capture, 130 m effective range, HDR + automated double-scan for transient-occlusion removal). It contains 100 equirectangular panoramas across 11 urban locations in Zürich, each paired with a dense metric depth map and a validity mask. It is used as the outdoor evaluation benchmark for PaGeR.

Dataset Summary

• Content: 100 outdoor scans across 11 Zürich locations, evaluation only.
• Modality: RGB (JPG), Depth (16-bit PNG, meters via scale factor), Validity Mask (8-bit PNG), Depth Viz (8-bit Spectral RGB PNG, preview only).
• Resolution: 4096 × 2048 equirectangular (ERP).
• Use Case: Evaluating long-range outdoor panoramic depth estimation.

Data Structure

A single test split with 100 rows, one per panorama. Each row carries:

Column	Type	Description
id	string	Sample id (<Location>- s<NNN>).
rgb	Image	8-bit equirectangular RGB (4096 × 2048, JPG-encoded).
depth	Image	16-bit single-channel PNG, (2048, 4096). Decode to meters as np.asarray(img, dtype=np.float32) * (200.0 / 65535.0). Invalid pixels are 0.0.
depth_viz	Image	8-bit RGB PNG, Spectral-colormapped log-depth (per-sample min/max stretch, median-filtered). Preview only — do NOT use for metrics or training; decode depth instead.
mask	Image	8-bit single-channel PNG, (2048, 4096). Decode as np.asarray(img, dtype=bool) (255 → True). True = reliable pixel; False = sky, no-return, or specular surface (glass façades). Always apply when computing depth metrics.

How to Use

import numpy as np
from datasets import load_dataset

ds = load_dataset("prs-eth/ZuriPano", split="test")
sample = ds[0]

rgb   = sample["rgb"]                                                       # PIL.Image, (W=4096, H=2048)
depth = np.asarray(sample["depth"], dtype=np.float32) * (200.0 / 65535.0)   # (2048, 4096) float32, meters
mask  = np.asarray(sample["mask"], dtype=bool)                              # (2048, 4096) bool

# Always apply the mask before computing depth metrics
valid_depth = depth[mask]

License

ZüriPano is released under the Creative Commons Attribution 4.0 International License (CC BY-4.0). You are free to share and adapt it for any purpose, including commercial use, as long as you attribute the PaGeR Authors and the ZüriPano dataset.

Citation

If you use ZüriPano in your work, please cite the PaGeR paper:

@article{bozic2026pager,
  title   = {Unified Panoramic Geometry Estimation via Multi-View Foundation Models},
  author  = {Bozic, Vukasin and Slavkovic, Isidora and Narnhofer, Dominik and
             Metzger, Nando and Rozumny, Denis and Schindler, Konrad and
             Kalischek, Nikolai},
  journal = {arXiv preprint arXiv:2605.26368},
  year    = {2026}
}