Research map / 3D vision / Depth and stereo
Dense-SfM: Structure from Motion with Dense Consistent Matching
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.
Dense-SfM builds a structure-from-motion pipeline on dense two-view matchers (RoMa/DKM), extending feature tracks via Gaussian Splatting visibility and refining them with a transformer + Gaussian Process multi-view kernelized matcher plus bundle adjustment.
Problem
Structure from Motion from pairwise dense matches yields fragmentary short feature tracks, limiting accuracy and point density, especially in texture-less regions.
Approach
Extend short two-view tracks by projecting triangulated points onto additional images whose visibility is checked with a Gaussian Splatting rendering, then refine the longer tracks with a multi-view kernelized matcher combining transformer features and a Gaussian Process positional embedding, with learned per-view confidence.
Key results
- On ETH3D triangulation at 1cm, RoMa + Ours reaches 84.79% accuracy and 7.38% completeness versus 79.32% / 3.13% for RoMa + DFSfM.
- On Texture-Poor SfM, RoMa + Ours achieves AUC@10 of 81.41 versus 69.01 for RoMa + DFSfM.
- Ablation shows removing GS track extension drops 1cm accuracy from 84.79% to 77.45%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add mutual verification to dense matches (ablation) | Dense-SfM with RoMa · ETH3D | accuracy (1cm): 84.03 -> 84.79 | Helped |
| add Gaussian Splatting track extension (ablation) | Dense-SfM with RoMa · ETH3D | accuracy (1cm): 77.45 -> 84.79 | Helped |
| replace quantized matching (r=4) with Gaussian Splatting track extension (ablation) | Dense-SfM with RoMa · ETH3D | accuracy (1cm): 82.82 -> 84.79 | Helped |
| use the full refinement module instead of the DFSfM multi-view transformer (ablation) | Dense-SfM with RoMa · ETH3D | accuracy (1cm): 80.59 -> 84.79 | Helped |
| add Gaussian Process to refinement (ablation) | Dense-SfM with RoMa · ETH3D | accuracy (1cm): 72.47 -> 84.79 | Helped |
| add the multi-view transformer to refinement (ablation) | Dense-SfM with RoMa · ETH3D | accuracy (1cm): 83.39 -> 84.79 | Helped |
| use reference-view confidence-based track selection (ablation) | Dense-SfM with RoMa · ETH3D | accuracy (1cm): 83.81 -> 84.79 | Helped |
| add iterative refinement (ablation) | Dense-SfM with RoMa · ETH3D | accuracy (1cm): 69.77 -> 84.79 | Helped |
| use 2 instead of 1 refinement iteration (ablation) | Dense-SfM with RoMa · ETH3D | accuracy (1cm): 81.12 -> 84.79 | Helped |
| use 3 instead of 2 refinement iterations (ablation) | Dense-SfM with RoMa · ETH3D | accuracy (1cm): 84.79 -> 85.01 | Helped |
| use 3 instead of 2 refinement iterations (ablation) | Dense-SfM with RoMa · ETH3D | completeness (1cm): 7.38 -> 6.74 | Hurt |
Surprising findings
- Mast3r-SfM baseline has very low 1cm triangulation accuracy (42.02%) despite its pose performance, while adding the proposed refinement raises it to 77.22%.
- Quantized matching with GS extension removed yields higher 1cm completeness (9.45%) than the full model (7.38%), a trade-off with accuracy.
Limitations
- GS-based track extension is challenging on IMC scenes with heavy occlusion and appearance change, so the quantization variant is used there instead.
- Accuracy and completeness trade off; some configurations gain completeness at the cost of accuracy.
Experimental setup
Datasets: MegaDepth ETH3D Texture-Poor SfM IMC Phototourism 2021
Benchmarks: ETH3D Texture-Poor SfM IMC 2021
Metrics: triangulation accuracy completeness AUC of pose error
Baselines: DFSfM PixSfM VGGSfM Mast3r-SfM LoFTR COLMAP AspanFormer MatchFormer SuperPoint+SuperGlue R2D2 D2-Net
Builds on: DFSfM DKM RoMa Gaussian Splatting SplaTAM PixSfM
Closest papers by meaning
- Detector-Free Structure from Motion (CVPR 2024 · similarity 0.88)
- SfM-Free 3D Gaussian Splatting via Hierarchical Training (CVPR 2025 · similarity 0.87)
- KeyGS: A Keyframe-Centric Gaussian Splatting Method for Monocular Image Sequences (AAAI 2025 · similarity 0.86)
- VGGSfM: Visual Geometry Grounded Deep Structure From Motion (CVPR 2024 · similarity 0.86)
- Light3R-SfM: Towards Feed-forward Structure-from-Motion (CVPR 2025 · similarity 0.86)
- DGS-LRM: Real-Time Deformable 3D Gaussian Reconstruction From Monocular Videos (NeurIPS 2025 · similarity 0.85)
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.