Long-Tail Internet Photo Reconstruction

TL;DR: We go beyond densely captured imagery to train more general 3D foundation models from the long tail of noisy, sparse, incomplete Internet photo collections of 3D scenes. We achieve this through MegaDepth-X, a large new dataset of scenes with high-quality 3D supervision, and a new way of simulating difficult image sets for training.

Abstract

Internet photo collections exhibit an extremely long-tailed distribution: a few famous landmarks are densely photographed and easily reconstructed in 3D, while most real-world sites are represented with sparse, noisy, uneven imagery beyond the capabilities of both classical and learned 3D methods. We believe that tackling this long-tail regime represents one of the next frontiers for 3D foundation models. Although reliable ground-truth 3D supervision from sparse scenes is challenging to acquire, we observe that it can be effectively simulated by sampling sparse subsets from well-reconstructed Internet landmarks. To this end, we introduce MegaDepth-X, a large dataset of 3D reconstrucions with clean, dense depth, together with a strategy for sampling sets of training images that mimic camera distributions in long-tail scenes. Finetuning 3D foundation models with these components yields robust reconstructions under extreme sparsity, and also enables more reliable reconstruction in symmetric and repetitive scenes, while preserving generalization to standard, dense 3D benchmark datasets.

The MegaDepth-X Dataset

Data preparation

Learning in the long-tail regime requires high-quality 3D supervision derived from Internet photo collections. Our data pipeline combines MASt3R-SfM together with Doppelgangers++, which reconstructs ambiguity-heavy scenes more reliably than COLMAP, especially under repetitive structures and challenging viewpoint distributions. We then run multi-view stereo (MVS) with COLMAP to obtain dense depth maps, followed by post-processing to further refine depth quality.

Filtering bad SFM reconstructions

We also manually verify reconstructed scenes against external references (e.g., Google Maps and Google Earth), and discard any scene that does not align with the corresponding bird's-eye-view geometry. All four examples shown here are bad SfM reconstructions: in (a), non-static human motions cause SfM to fail; in (b) and (c), inconsistent objects dominate the images; and in (d), reconstruction is degraded by doppelganger ambiguities.

After this filtering stage, we eventually curate MegaDepth-X, which is substantially larger than the commonly used in-the-wild reconstruction dataset, MegaDepth.

1865

MegaDepth-X

reconstructions

7x

larger

266

MegaDepth

reconstructions

Dense depth refinement

We refine MVS depth maps to filter out inaccurate depth estimates. In the first row (red box), our filtering mitigates depth bleeding i . In the second row, around the hanging light strings in the sky, our filtering removes bad depth caused by transient objects, producing cleaner depth than MegaDepth filtering.

Image

MegaDepth filtering

Our filtering

Simulating Long-Tail Scenes

Reliable long-tail supervision is scarce, so direct training in this regime is limited. We therefore simulate long-tail training batches by sampling sparse views from dense SfM reconstructions. Our sampling strategy aims to select views that span diverse viewpoints, maintain sparse overlap to emulate wide-baseline long-tail conditions, and still preserve sufficient covisibility within each scene component.

The panel below visualizes this sampling process with iterative greedy view sampling. For each training batch, we sample two disconnected view groups from a single scene graph to simulate two disjoint scene components, with eight views in each group. What's in the animation? i

<

Greedy step 1

> Auto-play

Current step

This animation visualizes iterative greedy view sampling.

Scene

Experimental Results

Acknowledgements

BibTeX

@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 Pattern Recognition},
  year={2026}
}