MoGe-2: Accurate Monocular Geometry with Metric Scale and Sharp Details

1 min read

MoGe-2 extends MoGe from scale-free point maps to metric ones without losing the relative accuracy of the affine-invariant representation. The second change is to the data: the authors found that noise and gaps in real training data blur fine detail, so they filter and complete real labels using sharp synthetic labels before training.

Findings

  • The abstract claims accurate relative geometry, precise metric scale and fine detail together, which it says no earlier method achieved at once. No headline number in the abstract.
  • Other tools reuse it as a scale source: InfiniDepth’s README loads a MoGe-2 checkpoint to recover metric scale for point cloud export.

Limits

  • I read the abstract and README only.
  • Metric scale from one image is a learned prior about object sizes. Underwater, where familiar objects are rare and haze changes apparent size cues, that prior is the part I would trust least.

Relevance

Metric depth from a single cheap camera is what a small vehicle without stereo would want. Whether the scale survives the domain shift is the question cai-2025-underwater-metric-depth tests for similar models, with poor results out of the box. See foundation models for robotics.