DfD-GSDefocus Image Prior for Metric-Scale Single-View Depth Estimation and View Synthesis

1 Keio University2 University of Stuttgart

An all-in-focus image and a defocused image feed a metric depth predictor, followed by a Gaussian predictor. The reconstructed piano scene illustrates metric-scale geometry.

Abstract

Recovering metric-scale geometry from observations at a single viewpoint remains fundamentally ill-posed, limiting the reliability of novel view synthesis. Recent methods rely on learned monocular depth priors, which provide strong structural estimates but can suffer from scale ambiguity and dataset-dependent biases. We propose DfD-GS, a feed-forward framework that predicts a metric-scale 3D Gaussian scene representation from a single-view all-in-focus–defocused image pair. The key idea is to exploit defocus blur as a physical metric depth cue. DfD-GS first estimates a defocus-based metric disparity map from the all-in-focus–defocused image pair and the corresponding defocus strength. This estimate is used to align and refine a monocular depth prior to metric scale. Visible and hidden 3D Gaussians are then predicted for occlusion-aware novel view synthesis. Experiments on synthetic-defocus datasets constructed from Booster and Middlebury show that DfD-GS outperforms single-image baselines in metric depth estimation and novel view synthesis. On WildRGB-D, DfD-GS achieves lower AbsRel and LPIPS but higher depth RMSE than the single-image baselines. Additional analyses show that DfD-GS maintains a more stable metric scale under field-of-view changes. Qualitative results on a real dataset show recovery of the main scene structure under optical defocus.

Method

DfD-GS pipeline: Depth Anything 3 features and disparity prior, defocus-conditioned U-Net, global alignment, local scale-map refinement, and visible and hidden Gaussian parameter prediction.

The method takes an all-in-focus–defocused image pair from the same viewpoint, with known camera intrinsics, defocus strength, and focus distance. The focus plane is placed at the nearest scene depth.

A frozen Depth Anything 3 backbone extracts multi-scale features and a relative geometry prior from the all-in-focus image. A defocus-conditioned U-Net estimates metric disparity offsets, which are used to globally align the prior. A learned scale map then refines the aligned disparity. The refined depth, input image, and DA3 features are used to predict visible and hidden 3D Gaussians for novel view synthesis.

Results

RGB-D Datasets with Synthetic Defocus

Booster / Metric depth estimation

Metric depth estimation comparison on Booster: input pair, ground truth, baseline predictions, and DfD-GS results. Figure brightness is increased by a factor of 1.2, as in the paper. Depth maps use the corresponding ground-truth depth range for visualization.
The all-in-focus–defocused input pair, ground-truth depth, and predictions from baseline methods and DfD-GS. Depth maps use the ground-truth depth range for visualization. Input brightness is increased by a factor of 1.2 for display.

Booster / Novel view synthesis

Novel view synthesis comparison on Booster: input pair, ground truth, baseline predictions, and DfD-GS results. Figure brightness is increased by a factor of 1.2, as in the paper.
The all-in-focus–defocused input pair, ground-truth target view, and views rendered by baseline methods and DfD-GS. Image brightness is increased by a factor of 1.2 for display.

Middlebury / Metric depth estimation

Metric depth estimation comparison on Middlebury: input pair, ground truth, baseline predictions, and DfD-GS results. Depth maps use the corresponding ground-truth depth range for visualization.
The all-in-focus–defocused input pair, ground-truth depth, and predictions from baseline methods and DfD-GS. Depth maps use the ground-truth depth range for visualization.

Middlebury / Novel view synthesis

Novel view synthesis comparison on Middlebury: input pair, ground truth, baseline predictions, and DfD-GS results.
The all-in-focus–defocused input pair, ground-truth target view, and views rendered by baseline methods and DfD-GS.

WildRGB-D / Metric depth estimation

Metric depth estimation comparison on WildRGB-D: input pair, ground truth, baseline predictions, and DfD-GS results. Depth maps use the corresponding ground-truth depth range for visualization.
The all-in-focus–defocused input pair, ground-truth depth, and predictions from baseline methods and DfD-GS. Depth maps use the ground-truth depth range for visualization.

WildRGB-D / Novel view synthesis

Novel view synthesis comparison on WildRGB-D: input pair, ground truth, baseline predictions, and DfD-GS results.
The all-in-focus–defocused input pair, ground-truth target view, and views rendered by baseline methods and DfD-GS.

Real Captures with Optical Defocus

Depth comparisons on five real scenes captured at 70 mm. Columns show the reference image, DA3 (metric), Depth Pro, Metric3D v2, MoGe, UniDepth v1, UniDepth v2, and DfD-GS. Labels show measured and estimated distances in cm. Depth maps are normalized for visualization.
Metric depth comparison on five real scenes captured at 70 mm. Reference shows the input images and measured distances; the other columns show estimated depth maps and distances in cm. Depth maps are normalized for visualization.

BibTeX

@misc{ueda2026dfdgs,
  title = {{DfD-GS}: Defocus Image Prior for Metric-Scale Single-View Depth Estimation and View Synthesis},
  author = {Ueda, Shiori and Yagawa, Rinto and Saito, Hideo and Mori, Shohei},
  year = {2026},
  eprint = {2611.26072},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.26072}
}