Refin3R: Geometry-Guided Refinement
for Feed-Forward 3D Reconstruction

Preprint 2026

KAIST AI

† Corresponding author

Teaser

Two indoor scenes: partial SfM reconstruction is refined into dense point maps, including regions without SfM coverage. Error maps compare DA3, SfM, GGPT, and Refin3R.

Feed-forward models such as Depth Anything 3 (DA3) produce dense but less accurate point maps. Refin3R uses SfM points to refine these dense predictions, improving accuracy even beyond SfM coverage. The error maps on the right compare DA3, SfM, GGPT, and Refin3R.

Method

Refin3R pipeline: feed-forward depth, geometry features, and partial SfM depth enter interacting 2D and 3D U-Nets. Spatial propagation fuses the estimates into refined depth and point maps.

Stage 1

Input configuration

We start with dense feed-forward predictions and accurate but incomplete SfM geometry.

Stage 2

Dual-branch refinement

Refin3R uses image context and SfM points from multiple views to improve the initial predictions.

Stage 3

Spatial fusion

The refined estimates are combined into dense, accurate 3D reconstructions.

Qualitative Results

Quantitative Results

Showing ScanNet++ results.

Contributions

  1. Dense and accurate reconstruction. Refin3R improves the accuracy of feed-forward predictions while preserving dense scene coverage.
  2. Effective across diverse scenes. Experiments demonstrate state-of-the-art reconstruction accuracy across diverse scenes.
  3. Efficient refinement. A fast refinement step improves reconstruction accuracy with little additional processing time.

BibTeX

@misc{lee2026refin3rgeometryguidedrefinementfeedforward,
      title={Refin3R: Geometry-Guided Refinement for Feed-Forward 3D Reconstruction},
      author={Jongmin Lee and Jaeyeong Kim and Seungryong Kim},
      year={2026},
      eprint={2611.23051},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2611.23051},
}

Figure

PDF ↗