MegaDepth-X

MegaDepth-X is a large-scale Internet photo reconstruction dataset for long-tail 3D reconstruction from Internet photos. For more details, please visit the project page.

Paper

Title: Long-tail Internet Photo Reconstruction
Authors: Yuan Li, Yuanbo Xiangli, Hadar Averbuch-Elor, Noah Snavely, Ruojin Cai
Paper: https://arxiv.org/abs/2604.22714
Project page: https://megadepth-x.github.io/

Abstract

Internet photo collections are highly long-tailed: a small number of famous landmarks have dense coverage and are easy to reconstruct, while most real-world scenes are sparse, noisy, and weakly connected. MegaDepth-X is introduced as a large dataset of clean 3D reconstructions with dense depth, together with a sparsity-aware sampling strategy that simulates long-tail camera distributions. This makes it possible to fine-tune 3D foundation models for robust reconstruction under extreme sparsity, symmetry, and ambiguity.

Release Structure

The public release is organized as scene archives under release/:

MegaDepth-X/
├── release/
│   ├── 30_Rockefeller_Plaza.tar.gz
│   ├── 875_North_Michigan_Avenue.tar.gz
│   ├── ...
├── license/
│   └── license.parquet
└── README.md

Each scene archive stores one or more numbered reconstruction components. A typical extracted layout is:

<scene>.tar.gz
└── 0/
    ├── images/
    │   ├── {image_name}
    │   └── ...
    ├── depths/
    │   ├── {image_name}.npz
    │   └── ...
    └── sparse/
        ├── cameras.bin
        ├── frames.bin
        ├── images.bin
        ├── points3D.bin
        └── rigs.bin

The directories are:

• images/: RGB images used by the reconstruction.
• depths/: Per-image compressed depth files in .npz format. Depth filenames match the corresponding image filenames.
• sparse/: Sparse reconstruction files in COLMAP binary format.

Licensing Metadata

The MegaDepth-X dataset assets, including depth maps, SfM models, and related reconstruction files, are licensed under the Creative Commons Attribution 4.0 International License. The original images come with their own licenses.

Following the MegaScenes dataset, we also provide a table that indexes the released MegaDepth-X images in license/license.parquet. Parquet stores tabular data like CSV, but is more compact and faster to read. It can be read using Python dataframe libraries such as Polars or Pandas.

Each row corresponds to one released image. The columns are:

• name_in_megadepth-x (str): Relative path of the image inside MegaDepth-X, for example <scene>/<component>/images/<image_file>.
• cat (str): Scene name.
• subcat (str): Upstream Wikimedia Commons subcategory associated with the image.
• image (str): Relative path used in MegaScenes.
• image_name (str): Parsed image filename.
• license_id (str): Parsed license identifier for the image.
• license_url (str): License URL for the image.
• license_short_name (str): Abbreviated license name for the image.
• usage_terms (str): Parsed usage terms or license description for the image.

While this table contains parsed licensing information derived from upstream metadata, we have proactively removed specific scenes and images that no longer align with open-access or research-friendly licenses. We encourage users to verify image licenses themselves before further redistribution.

Note: Some reconstruction components may have missing images compared to their initial SfM logs; this is intentional to comply with licensing restrictions of the original image assets.

Citation

If you use MegaDepth-X in your research, please cite:

@inproceedings{li2026longtail,
  title={Long-Tail Internet Photo Reconstruction},
  author={Li, Yuan and Xiangli, Yuanbo and Averbuch-Elor, Hadar and Snavely, Noah and Cai, Ruojin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Patter Recognition},
  year={2026}
}

@inproceedings{tung2024megascenes,
  title={MegaScenes: Scene-Level View Synthesis at Scale}, 
  author={Tung, Joseph and Chou, Gene and Cai, Ruojin and Yang, Guandao and Zhang, Kai and Wetzstein, Gordon and Hariharan, Bharath and Snavely, Noah},
  booktitle={Proceedings of the European Conference on Computer Vision},
  year={2024}
}