Research map / 3D vision / Depth and stereo

MP-SfM: Monocular Surface Priors for Robust Structure-from-Motion

CVPR 2025 · 3D vision › Depth and stereo · approach: Feed-forward 3D reconstruction · 7 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.

Augments incremental SfM (COLMAP-style) with monocular depth and normal priors from pretrained networks, enabling registration from two-view tracks and rejecting symmetry-induced false registrations via a dense depth consistency check.

Problem

Structure-from-Motion fails under extreme viewpoint changes (low overlap), low parallax, and repeated/symmetric structure because classical pipelines require three-view tracks to fix scale and cannot reject faulty symmetric associations.

Approach

Lift single-view points using monocular depth so images can be registered without three-view overlap; jointly refine poses, points, and depth maps via alternating optimization that conditions depth on monocular depth and normal priors with uncertainty propagation; and use a dense forward-backward depth consistency check to reject incorrectly posed views.

Key results

• On ETH3D minimal three-view triplets, the SP+LG ours variant reaches 27.3/55.9/71.8 AUC@1/5/20° versus 8.4/15.8/22.5 for SP+LG GLOMAP.
• On SMERF minimal overlap, ours (SP+LG) achieves 9.2/41.0/69.8 AUC while COLMAP, SLR and GLOMAP are at or below 3.5/8.0/13.3.
• On RealEstate10k low-parallax, MASt3R ours reaches 35.5/81.9/91.5 AUC@1/10/30° versus 33.4/80.8/91.2 for MASt3R-SfM.

Measured effects reported in the paper

Surprising findings

• Depth refinement with surface normals and depth regularization can hurt accuracy in well-constrained high-overlap scenes, since monocular priors are noisier than reprojection errors.
• Estimating depth uncertainty is critical: Metric3D-v2 without depth uncertainty drops substantially, and models lacking uncertainty (Depth Pro, Depth Anything v2) underperform.

Limitations

• Relies on reliable uncertainty estimates for monocular priors, which many depth models lack or are overconfident about.
• Normal estimation is difficult for some surfaces such as vegetation, limiting outdoor performance.
• Increases SfM runtime and still depends on image retrieval and matching components that fail in extreme conditions.

Experimental setup

Datasets: ETH3D Tanks & Temples SMERF RealEstate10k

Benchmarks: ETH3D Tanks & Temples SMERF RealEstate10k

Metrics: AUC of relative pose error at 1/5/20 degrees AUC at 1/10/30 degrees

Baselines: COLMAP GLOMAP MASt3R-SfM DF-SfM VGG-SfM Structure-less resectioning StudioSfM Doppelgangers

Builds on: COLMAP Metric3D v2 DSINE SuperPoint+LightGlue StudioSfM

Closest papers by meaning

• VGGSfM: Visual Geometry Grounded Deep Structure From Motion (CVPR 2024 · similarity 0.88)
• Multi-view Reconstruction via SfM-guided Monocular Depth Estimation (CVPR 2025 · similarity 0.87)
• Light3R-SfM: Towards Feed-forward Structure-from-Motion (CVPR 2025 · similarity 0.87)
• MonoPhysics: Estimating Geometry, Appearance, and Physical Parameters from Monocular Videos (NeurIPS 2026 · similarity 0.86)
• Zero-Shot Scene Reconstruction from Single Images with Deep Prior Assembly (NeurIPS 2024 · similarity 0.86)
• DiffusionSfM: Predicting Structure and Motion via Ray Origin and Endpoint Diffusion (CVPR 2025 · similarity 0.86)