cornell-vailab/megaunscene

<div align="center"> <h1>Emergent Extreme-View Geometry in 3D Foundation Models</h1>

<a href="https://arxiv.org/abs/2511.22686"><img src="https://img.shields.io/badge/arXiv-2511.22686-b31b1b" alt="arXiv"></a> &nbsp; <a href="https://cornell-vailab.github.io/Ext-3DFMs/"><img src="https://img.shields.io/badge/ProjectPage-green" alt="Project Page"></a> &nbsp; <a href="https://github.com/theREALevan/extreme-view-3dfm"><img src="https://img.shields.io/badge/MainCode-black?logo=github" alt="Main Code Repository"></a> &nbsp; <a href="https://github.com/jot-jt/extreme-view-3dfm-gen-eval"><img src="https://img.shields.io/badge/Generalization_Evaluations-black?logo=github" alt="Generalization Eval Repository"></a> &nbsp; <a href="https://huggingface.co/datasets/cornell-vailab/megaunscene"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Dataset-blue" alt="Hugging Face"></a>

Yiwen Zhang¹ &nbsp; Joseph Tung² &nbsp; Ruojin Cai³ &nbsp; David Fouhey² &nbsp; Hadar Averbuch-Elor¹

¹Cornell University &nbsp; ²New York University &nbsp; ³Kempner Institute, Harvard University </div>

MegaUnScene Benchmark

Overview

MegaUnScene is a dataset of Internet scenes unseen by existing 3DFMs for benchmarking. There are three test splits split across two evaluation tasks:

• —Relative Pose Estimation: UnScenePairs and UnScenePairs-t
• —Dense 3D Reconstruction: UnSceneRecon

Benchmarking Relative Pose in the Wild

UnScenePairs targets image pairs with predominant rotational motion, while UnScenePairs-t focuses on pairs with larger camera baselines. Unlike prior benchmarks, these subsets capture unconstrained, in-the-wild views unseen by 3DFMs. In total, they comprise over 6,000 image pairs across more than 450 scenes, including substantial non-overlapping splits.

UnSceneRecon: Benchmarking Dense Reconstruction in the Wild

UnSceneRecon is a subset comprising 100 in-the-wild reconstructions with metric scale annotations. This benchmark evaluates dense reconstruction quality on unconstrained Internet photos exhibiting diverse lighting conditions, transients, and varying camera models.

Dataset Download

Please refer to https://huggingface.co/docs/hub/en/datasets-downloading on how to download datasets from HuggingFace.

We provide a post-processing script to prepare the dataset. In the direct download, depth maps are compressed with bit shuffling to minimize download size. Each scenes' image and depth map folders are zipped. This script unzips all folders and decompresses all depth maps in a given base directory.

Example use: python unzip_and_decompress_megaunscene.py --megaunscene_base {PATH_TO_MEGAUNSCENE_ROOT}

Note: Python library numcodecs is required to decompress depth maps; it is installable with pip install numcodecs.

Dataset Size

The full compressed dataset download is ~202G, or ~706G uncompressed.

Dataset Structure

megaunscene/
├── images.csv                          # Metadata for all images
├── reconstructions.csv                 # Metadata for all reconstructions
└── scenes/                             # Main scene data directory
    ├── {scene_name}/                   # Scene name
    │   ├── {recon_id}/                 # Reconstruction ID
    │   │   ├── images/
    │   │   │   └── ...
    │   │   ├── depth_maps/
    │   │   │   └── ...
    │   │   └── sparse/                 # Sparse reconstruction data (COLMAP format)
    │   │       ├── cameras.bin
    │   │       ├── images.bin
    │   │       ├── points3D.bin
    │   │       └── ...
    │   └── ... (additional reconstructions)
    └── ... (469 scenes total)

Metadata

reconstructions.csv

reconstructions.csv contains metadata for all reconstructions in the dataset.

images.csv

images.csv contains metadata for all images in the dataset, including licensing information sourced from Wikimedia Commons.

Notes

• —Scene names: URL-encoded (spaces → underscores, slashes → %2F)
• —Image paths: Relative to the directory in the images/ folder
• —File locations:
• —Images: scenes/{scene}/{recon_id}/images/{image_name}, where image_name is full relative path defined in images.csv
• —Depth maps: scenes/{scene}/{recon_id}/depth_maps/{image_name}.npy (if has_depth=true)
• —Sparse reconstruction: scenes/{scene}/{recon_id}/sparse/

Citation

If you found this dataset helpful, please cite

@misc{zhang2025emergentextremeviewgeometry3d,
      title={Emergent Extreme-View Geometry in 3D Foundation Models}, 
      author={Yiwen Zhang and Joseph Tung and Ruojin Cai and David Fouhey and Hadar Averbuch-Elor},
      year={2025},
      eprint={2511.22686},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.22686}, 
}

License

This dataset is licensed under the Creative Commons Attribution 4.0 International License. The photos for each scene are sourced from Wikimedia Commons and have their own licenses; please see images.csv for additional details.