prs-eth/PanoInfinigen
收藏资源简介：
--- license: - bsd-3-clause - cc-by-nc-4.0 task_categories: - depth-estimation - image-to-image tags: - panorama - depth - normals - synthetic - equirectangular - '360' - infinigen - icity pretty_name: PanoInfinigen configs: - config_name: indoor data_files: - split: train path: "data/indoor/train-*" - split: val path: "data/indoor/val-*" - split: test path: "data/indoor/test-*" - config_name: nature data_files: - split: train path: "data/nature/train-*" - split: val path: "data/nature/val-*" - split: test path: "data/nature/test-*" - config_name: urban data_files: - split: train path: "data/urban/train-*" - split: val path: "data/urban/val-*" - split: test path: "data/urban/test-*" dataset_info: - config_name: indoor features: - name: image dtype: image - name: depth dtype: binary - name: normals dtype: binary - config_name: nature features: - name: image dtype: image - name: depth dtype: binary - name: normals dtype: binary - config_name: urban features: - name: image dtype: image - name: depth dtype: binary - name: normals dtype: binary --- <h1 align="center" style="font-size: 2.8em; margin: 0.4em 0;">🗃️ PanoInfinigen Dataset</h1> <p align="center"> <a title="Github" href="https://github.com/prs-eth/PaGeR" target="_blank" rel="noopener noreferrer" style="display: inline-block;"> <img src="https://img.shields.io/github/stars/prs-eth/PaGeR?label=GitHub%20%E2%98%85&logo=github&color=C8C" alt="Github"> </a> <a title="Website" href="https://pager360.github.io/" target="_blank" rel="noopener noreferrer" style="display: inline-block;"> <img src="https://img.shields.io/badge/%E2%99%A5%20Project%20-Website-blue" alt="Website"> </a> <a title="arXiv" href="https://arxiv.org/abs/2605.26368" target="_blank" rel="noopener noreferrer" style="display: inline-block;"> <img src="https://img.shields.io/badge/%F0%9F%93%84%20Read%20-Paper-AF3436" alt="arXiv"> </a> <a title="Hugging Face Collection" href="https://huggingface.co/collections/prs-eth/pager-697241d06b3733a6f18e4d39" target="_blank" rel="noopener noreferrer" style="display: inline-block;"> <img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Collection-FFD21E" alt="Hugging Face Collection"> </a> <a title="License (indoor + nature)" href="https://opensource.org/licenses/BSD-3-Clause" target="_blank" rel="noopener noreferrer" style="display: inline-block;"> <img src="https://img.shields.io/badge/indoor%20%2B%20nature-BSD_3--Clause-blue.svg" alt="BSD-3-Clause"> </a> <a title="License (urban)" href="https://creativecommons.org/licenses/by-nc/4.0/" target="_blank" rel="noopener noreferrer" style="display: inline-block;"> <img src="https://img.shields.io/badge/urban-CC_BY--NC_4.0-yellowgreen.svg" alt="CC BY-NC 4.0"> </a> </p> **PanoInfinigen** is a synthetic dataset of high-resolution panoramic images in [ERP](https://en.wikipedia.org/wiki/Equirectangular_projection), featuring perfectly aligned RGB, Depth, and Surface Normals. This dataset was generated using a modified [Infinigen](https://infinigen.org/) framework to support wide-angle panoramic geometry, plus the [iCity](https://superhivemarket.com/products/icity) procedural city generator for the urban split. It serves as the primary training data for [PaGeR](https://pager360.github.io/), 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. ```python 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](https://infinigen.org/) (BSD-3-Clause) | **BSD-3-Clause** | | `urban` | Rendered from scenes generated with the [iCity](https://superhivemarket.com/products/icity) Blender addon | **CC BY-NC 4.0** (academic / non-commercial only) | ## Acknowledgements PanoInfinigen is built on top of [Infinigen](https://infinigen.org/) — 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](https://superhivemarket.com/products/icity) Blender addon. We are sincerely grateful to **Hothifa Smair** and the [**Parametra**](https://parametra.net/) 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](https://superhivemarket.com/products/icity) for the upstream license terms. ## Citation If you use PanoInfinigen in your work, please cite the PaGeR paper: ```bibtex @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} } ```
PanoInfinigen is a synthetic dataset of high-resolution panoramic images in Equirectangular Projection (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. The dataset is split into three configurations: indoor, nature, and urban, each containing train, validation, and test splits. Features include: image (8-bit RGB panoramic image, PNG format), depth (binary .npy file, float16 type, range [0, 75] m for indoor/nature, [0, 500] m for urban), and normals (binary .npy file, float16 type, range [-1, 1]). It is licensed under BSD-3-Clause for indoor and nature splits, and CC BY-NC 4.0 for urban split (academic/non-commercial only).
