Research map / 3D vision / Depth and stereo
MoGe-2: Accurate Monocular Geometry with Metric Scale and Sharp Details
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Extends the affine-invariant MoGe point-map model to metric geometry by decoupling a CLS-token-predicted global scale, and refines noisy real training depth with synthetic-model-based mismatch filtering and Poisson completion to recover sharp details.
Problem
Monocular geometry estimation lacks a single model that simultaneously gives accurate relative geometry, metric scale, and fine-grained boundary detail; noisy and misaligned real depth labels blur sharp structures.
Approach
Keep MoGe's affine-invariant point map branch and add a separate CLS-token-conditioned MLP that predicts a global metric scale with stop-gradient supervision, so scale does not disturb relative geometry. Real depth labels are filtered for local mismatches using a synthetic-only MoGe model's predictions, then holes are filled by log-space Poisson completion.
Key results
- On metric geometry over 7 datasets, the model reaches point Rel 8.19 and delta_1 93.6 versus MoGe-free baselines such as UniDepth V2 at 10.1 and 91.9, and Depth Pro at 13.7 and 81.9.
- On relative point-map evaluation averaged over 10 datasets, the model reports Rel 10.8 and delta_1 88.5, comparable to MoGe (7.46, 94.1) and better than all other baselines on the listed metrics.
- Boundary F1 averaged over iBims-1, HAMMER, Sintel and Spring is best in average rank 1.75, with iBims-1 F1 17.9 versus Depth Pro 14.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| decouple metric scale from relative geometry using a CLS-token-conditioned MLP instead of entangled SI-Log prediction (ablation) | MoGe-2 · 10-dataset ablation average | metric point Relp: 10.0 -> 9.20 | Helped |
| predict scale with a CLS-token-conditioned MLP rather than a convolutional head (ablation) | MoGe-2 · 10-dataset ablation average | metric depth Rel d: 17.7 -> 16.5 | Helped |
| use a CLS-token-conditioned MLP scale head rather than a convolutional scale head (ablation) | MoGe-2 · 10-dataset ablation average | boundary F1: 12.7 -> 12.5 | No clear effect |
| add real-world training data to synthetic-only training (ablation) | MoGe-2 · 10-dataset ablation average | metric depth Rel d: 21.7 -> 15.8 | Helped |
| add real-world training data to synthetic-only training (ablation) | MoGe-2 · 10-dataset ablation average | boundary F1: 13.3 -> 10.3 | Hurt |
| refine real-world training data rather than use raw real-world data (ablation) | MoGe-2 · 10-dataset ablation average | boundary F1: 10.3 -> 12.5 | Helped |
| refine real-world training data rather than use raw real-world data (ablation) | MoGe-2 · 10-dataset ablation average | metric depth Rel d: 15.8 -> 16.5 | Hurt |
| use MoGe-2 instead of MoGe | monocular geometry estimation · 10-dataset average relative geometry | average rank: 2.53 -> 2.05 | Helped |
| use MoGe-2 instead of UniDepth V2 | metric geometry estimation · 7-dataset average metric geometry | average rank: 2.51 -> 1.95 | Helped |
| use MoGe-2 instead of Depth Pro | boundary sharpness prediction · iBims-1, HAMMER, Sintel, Spring | average rank: 1.50 -> 1.75 | Hurt |
Surprising findings
- Training only on synthetic data gives the sharpest boundaries but markedly worse geometry accuracy, and a naive convolutional scale head worsens metric scale results relative to the CLS-token MLP.
Limitations
- Struggles with extremely fine structures such as thin lines and hair.
- Maintaining straight and aligned structures under large foreground-background scale differences is difficult.
- Metric scale ambiguity can cause deviations in out-of-distribution scenes.
Experimental setup
Datasets: A2D2 Argoverse2 ARKitScenes BlendedMVS MegaDepth ScanNet++ Taskonomy Waymo Hypersim TartanAir Structured3D IRS
Benchmarks: NYUv2 KITTI ETH3D iBims-1 GSO Sintel DDAD DIODE Spring HAMMER
Metrics: relative error (Rel) delta_1 inlier percentage boundary F1 average rank
Baselines: MoGe Depth Pro UniDepth V2 UniDepth V1 Metric3D V2 Depth Anything V2 MASt3R ZoeDepth
Builds on: MoGe DINOv2 Depth Anything V2 Depth Pro
Closest papers by meaning
- MoGe: Unlocking Accurate Monocular Geometry Estimation for Open-Domain Images with Optimal Training Supervision (CVPR 2025 · similarity 0.95)
- SurGe: Improved Surface Geometry in Point Maps (NeurIPS 2026 · similarity 0.90)
- Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric Refinement (NeurIPS 2026 · similarity 0.87)
- UniK3D: Universal Camera Monocular 3D Estimation (CVPR 2025 · similarity 0.85)
- UniDepth: Universal Monocular Metric Depth Estimation (CVPR 2024 · similarity 0.85)
- Feat2GS: Probing Visual Foundation Models with Gaussian Splatting (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.