DGSfM: Depth-Guided Scale-Aware Global Structure-from-Motion

2026-07-10 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

DGSfM, a novel depth-aware global Structure-from-Motion (SfM) pipeline, addresses the challenges of recovering camera poses and sparse 3D structure from unordered images. Traditional global SfM suffers from sensitivity to noisy baseline estimates, weak view-graph constraints, and false edges due to scale-ambiguous epipolar geometry. DGSfM integrates monocular depth maps as a scalable prior, converting scale-ambiguous epipolar constraints into scale-aware relative pose constraints using a depth-aware solver. It enhances robustness through view-graph filtering and depth-consistency-based correspondence pruning, suppressing false matches. Furthermore, global scale averaging and depth-guided pose-point initialization align monocular depth maps and provide stable initialization for global positioning and bundle adjustment. Experiments on ETH3D and IMC2021 demonstrate DGSfM's consistent improvement over strong global SfM baselines, yielding substantial gains in pose accuracy.

Key takeaway

For Computer Vision Engineers developing 3D reconstruction systems, DGSfM offers a robust approach to overcome the limitations of traditional global Structure-from-Motion. By incorporating monocular depth priors and advanced filtering, your pipelines can achieve significantly higher pose accuracy and resilience to noisy data. Consider integrating depth-aware relative pose solvers and depth-consistency pruning to enhance your SfM workflows, especially when dealing with challenging unordered image datasets.

Key insights

DGSfM enhances global SfM robustness and accuracy by integrating monocular depth maps and depth-aware filtering.

Principles

• Global SfM's reliance on scale-ambiguous epipolar geometry is a key challenge.
• Monocular depth maps can serve as a scalable prior for 3D reconstruction.
• Depth-consistency improves robustness against false matches and weak view-graph constraints.

Method

A depth-aware relative pose solver converts epipolar constraints to scale-aware ones. View-graph filtering and depth-consistency prune false edges/matches. Global scale averaging and depth-guided initialization align depth maps.

In practice

• Integrate monocular depth for improved SfM robustness.
• Apply depth-consistency checks to filter correspondences.
• Utilize depth-guided initialization for stable global positioning.

Topics

• Structure-from-Motion
• 3D Reconstruction
• Monocular Depth Estimation
• Camera Pose Estimation
• Computer Vision
• Global SfM

Code references

• sithu31296/DGSfM

Best for: Research Scientist, AI Scientist, Computer Vision Engineer

Related on AIssential

• Large Depth Completion Model from Sparse Observations — LDCM achieves robust, metric-accurate dense depth completion from sparse observations via a transformer, novel initialization, and 3D coordinate regression.
• 4DGS360: 360{\deg} Gaussian Reconstruction of Dynamic Objects from a Single Video — 4DGS360 enables consistent 360-degree dynamic object reconstruction from monocular video using 3D-native, occlusion-aware initialization.
• Frequency-Structured Field Learning for Light-Field Disparity Estimation — FreqLF combines global Fourier and local convolutional updates within a latent field for accurate light-field disparity estimation.
• FoundationGeo: Learning Spatial Pixel-Wise Fields for Monocular Metric Geometry — FoundationGeo improves monocular metric geometry via a two-stage framework, pixel-wise calibration, and intrinsic-aware data synthesis for robust zero-shot generalization.
• 🌔Foundation Global SFM🌔 👉Glob3R is a global SfM-style reconstruction built on 3D... — Glob3R uses 3D foundation models to optimize feed-forward geometric predictions for global SfM reconstruction.
• Visual Relocalization from Sparse Views in Aliased and Low-Texture Environments via Novel View Synthesis — Combining MVS and LiDAR depths in 3DGS training significantly improves visual relocalization accuracy in extreme environments.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.