CVPR 2025

Dense-SfM: Structure from Motion with
Dense Consistent Matching

Jongmin Lee1   Sungjoo Yoo1
1Seoul National University
Paper Code (will be open)

Summary

Dense-SfM overview

We present Dense-SfM, a novel SfM framework that integrates dense feature matching for accurate and dense 3D reconstruction from multi-view images. To resolve the track fragmentation problem inherent in pairwise dense matching, Dense-SfM leverages Gaussian Splatting to reason about 3D point visibility and extend short tracks to additional views — without sacrificing subpixel accuracy. The extended tracks are then refined by a novel multi-view kernelized matching module combining transformer and Gaussian Process architectures, followed by iterative bundle adjustment.

Method

Method overview

Stage 1

Initial SfM via dense matching

Dense pairwise matches (DKM / RoMa) filtered by mutual bidirectional verification build an initial SfM with short, mostly two-view tracks.

Stage 2

Track extension via Gaussian Splatting

3D points are treated as small Gaussians. After GS optimization, visibility-based projection extends each track to additional co-visible images.

Stage 3

Iterative SfM refinement

A multi-view kernelized matching module (transformer + Gaussian Process) refines track coordinates, followed by geometric bundle adjustment.

Contributions

1
Dense SfM framework — optimized for dense matchers, achieving accurate and dense 3D reconstruction even in texture-poor scenes.
2
GS-based track extension — visibility-driven track lengthening that preserves subpixel accuracy, outperforming quantization-based approaches.
3
Multi-view kernelized refinement — transformer and Gaussian Process architectures jointly leverage feature and positional information for precise track refinement.
4
State-of-the-art results — ETH3D, Texture-Poor SfM, and IMC PhotoTourism benchmarks, consistently outperforming detector-based and semi-dense baselines.

Visualization

Visualization results

Left: Qualitative Results on Texture-Poor SfM dataset.
Right: Qualitative Results on IMC 2021 dataset.

Related Links

There are excellent works that are related with ours and gave us inspiration.

MV-RoMa extends dense pairwise matching into a multi-view setting, jointly estimating dense correspondences from a source image to multiple co-visible targets in a single forward pass — enabling geometrically consistent track reconstruction for SfM.

Detector-Free SfM proposes a detector-free structure from motion framework that eliminates the requirement of keypoint detection and can recover poses even on challenging texture-poor scenes.

DKM and RoMa propose a significantly robust dense feature matcher.

BibTeX

@misc{lee2025densesfmstructuremotiondense,
      title={Dense-SfM: Structure from Motion with Dense Consistent Matching},
      author={JongMin Lee and Sungjoo Yoo},
      year={2025},
      eprint={2501.14277},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2501.14277},
}