Yuanbo Xiangli, Hanyu Chen, Xueqing Tsang, Noah Snavely
[Project Page, Dataset, Model Weights (WildMoGe)]
MetricScenes aggregates scenes from MegaScenes, Aerial-MegaDepth, and Stereo4D into a single metric-scale training corpus. In MetricScenes, for each image we provide RGB, partial depth from SfM/MVS or geometric priors, completed depth from our two-stage Poisson method, and camera metadata. Depth PNGs are 16-bit log-encoded; see [depth_completion/README.md](depth_completion/README.md#encoded-depth-png) for the decoding format.
This repository contains the implementation of: (1) our two-stage Poisson completion method that fuses MVS depth maps (background) with model predicted depth maps (foreground) to create complete depth maps (2) image registration pipeline adapted from MASt3R-SfM, enhanced with Doppelgangers++ classifier. We use this to register georeferenced street-level views to SfM models to scale the geometry to physical dimensions:
| Directory | Description |
|---|---|
| depth_completion | Edge-aware two-stage Poisson depth completion |
| mast3r_sfm | Image registration pipeline adapted from MASt3R-SfM w/ DG++ |
See each subdirectory's README for installation, dependencies, and usage.
WildMoGe is a MoGe-2 ViT-Large-Normal model fine-tuned on our MetricScenes dataset. For inference and evaluation, please follow MoGe.
The MetricScenes dataset is released under CC BY 4.0. Original source images retain their own licenses.
Component code may carry additional upstream licenses (e.g. MASt3R, DUSt3R, Doppelgangers++); see the respective subdirectories and upstream repositories.
@inproceedings{xiangli2026honey,
title={{Honey, I Shrunk the Arc de Triomphe!}},
author={Xiangli, Yuanbo and Chen, Hanyu and Tsang, Xueqing and Snavely, Noah},
booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
year={2026}
}