🗃️ PanoInfinigen Dataset

PanoInfinigen is a synthetic dataset of high-resolution panoramic images in ERP, featuring perfectly aligned RGB, Depth, and Surface Normals. This dataset was generated using a modified Infinigen framework to support wide-angle panoramic geometry, plus the iCity procedural city generator for the urban split.

It serves as the primary training data for PaGeR, a single-step diffusion model for zero-shot panoramic depth and normal estimation.

Dataset Summary

• Content: Synthetic indoor, nature, and urban scenes.
• Modality: RGB (PNG), Depth (binary .npy), Surface Normals (binary .npy).
• Projection: Equirectangular (ERP).
• Use Case: Training and evaluating monocular panoramic depth and normal estimation models.

Data Structure

The dataset is split into three configurations: indoor, nature, and urban. Each contains train, validation, and test splits.

Feature	Type	Description
image	PIL.Image	8-bit RGB Panoramic Image.
depth	binary	float16 NumPy array. Range: [0, 75] m for indoor / nature, [0, 500] m for urban.
normals	binary	float16 NumPy array. Range: [-1, 1].

How to Use

Since depth and normals are stored as binary blobs to preserve precision (float16), you need to use io.BytesIO to load them back into NumPy.

import io
import numpy as np
from datasets import load_dataset

# Load the indoor training split (swap name for "nature" or "urban")
ds = load_dataset("prs-eth/PanoInfinigen", name="indoor", split="train")

sample = ds[0]

# 1. Get RGB Image
rgb = sample["image"]

# 2. Convert Binary Depth to NumPy (float16)
#    Max range: 75 m (indoor / nature), 500 m (urban)
depth = np.load(io.BytesIO(sample["depth"]))

# 3. Convert Binary Normals to NumPy (float16, -1 to 1)
normals = np.load(io.BytesIO(sample["normals"]))

License

PanoInfinigen ships with two licenses, one per provenance:

Config(s)	Provenance	License
indoor, nature	Rendered from Infinigen (BSD-3-Clause)	BSD-3-Clause
urban	Rendered from scenes generated with the iCity Blender addon	CC BY-NC 4.0 (academic / non-commercial only)

Acknowledgements

PanoInfinigen is built on top of Infinigen — the indoor and nature splits were rendered with a modified Infinigen pipeline, and we thank the Infinigen team for releasing the upstream procedural generator under an open license.

The urban split was rendered from scenes generated with the iCity Blender addon. We are sincerely grateful to Hothifa Smair and the Parametra team for kindly granting us written authorization to use the iCity asset library to generate this data, to train PaGeR on it, and to release the resulting urban split for non-commercial academic research. This work would not have been possible without their generosity. If you use the urban split, please credit iCity, Hothifa Smair, and the Parametra team alongside the dataset citation, and consult the iCity product page for the upstream license terms.

Citation

If you use PanoInfinigen 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}
}