Research map / 3D vision / Depth and stereo

MoGe: Unlocking Accurate Monocular Geometry Estimation for Open-Domain Images with Optimal Training Supervision

CVPR 2025 · 3D vision › Depth and stereo · approach: Foundation models for monocular geometry · 69 citations (OpenAlex)

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.

Trains a ViT-based model (MoGe) to predict affine-invariant 3D point maps from single open-domain images, using a robust optimal efficient (ROE) alignment solver and multi-scale local geometry loss, yielding state-of-the-art point map, depth, and FOV estimation.

Problem

Monocular geometry estimation for open-domain images is ill-posed; scale, shift and focal-length ambiguity make supervision of 3D point maps from single images difficult and existing global alignment is suboptimal or outlier-sensitive.

Approach

Predict an affine-invariant point map (agnostic to global scale and 3D shift) to remove focal-distance ambiguity from supervision, paired with an exact, outlier-robust alignment solver (ROE) for the global loss and a multi-scale sphere-based local geometry loss with independent per-region alignment.

Key results

• Zero-shot on eight datasets, affine-invariant point map average Relp 1.31 vs 2.94 for DUSt3R and 1.81 for UniDepth, reducing error over 35% versus prior best point-cloud methods.
• Local point map region-wise Relp reduced from 7.97 to 5.50 (about 30% over previous methods).
• Camera FOV evaluation averages mean error 2.91 degrees and median 2.21 degrees, beating Perspective Fields and WildCamera-style baselines and point map methods.

Measured effects reported in the paper

Surprising findings

• Affine-invariant depth representation alone (with shift scalar and ray map) slightly improves some metrics but other representation variants drop dramatically in ablation.
• Assigning a large value (1000x average distance) to infinity/sky regions instead of a mask compromises foreground accuracy.

Limitations

• Relies on sky/infinity masks from SegFormer for real outdoor data, which adds dependency on external segmentation.
• Truncation of residuals makes the alignment objective non-convex, handled only with minor adjustments.
• Training uses ~9M frames with varied label quality, requiring hand-tailored loss combinations per dataset.

Experimental setup

Datasets: NYUv2 KITTI ETH3D iBims-1 Sintel GSO DDAD DIODE 21 publicly available datasets (~9M frames) for training

Benchmarks: NYUv2 KITTI ETH3D iBims-1 GSO Sintel DDAD DIODE

Metrics: Relp delta_p 1 Reld delta_d 1 camera FOV mean error camera FOV median error

Baselines: LeReS DUSt3R UniDepth Metric3D V2 ZoeDepth Depth Anything V1 Depth Anything V2 Marigold GeoWizard MiDaS V3.1 Perspective Fields WildCamera

Builds on: DINOv2 DUSt3R UniDepth Depth Anything V2 Marigold

Closest papers by meaning

• MoGe-2: Accurate Monocular Geometry with Metric Scale and Sharp Details (NeurIPS 2025 · similarity 0.95)
• SurGe: Improved Surface Geometry in Point Maps (NeurIPS 2026 · similarity 0.88)
• Geometry-Guided Domain Generalization for Monocular 3D Object Detection (AAAI 2024 · similarity 0.86)
• Dens3R: A Foundation Model for 3D Geometry Prediction (ICLR 2026 · similarity 0.86)
• MonoInstance: Enhancing Monocular Priors via Multi-view Instance Alignment for Neural Rendering and Reconstruction (CVPR 2025 · similarity 0.86)
• MonoMAE: Enhancing Monocular 3D Detection through Depth-Aware Masked Autoencoders (NeurIPS 2024 · similarity 0.85)