Research map / 3D vision / Depth and stereo
DualFocus: Depth from Focus with Spatio-Focal Dual Variational Constraints
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.
Proposes DualFocus, a depth-from-focus network that adds spatial variational constraints (integrable gradient surface reconstruction weighted by sharpness) and focal variational constraints (unimodal, monotonic focus probabilities) to reach state-of-the-art on NYU Depth v2, FoD500, and DDFF 12-Scene.
Problem
Depth-from-focus struggles in complex scenes with fine textures or abrupt depth changes, where focus cues are ambiguous or misleading.
Approach
Predict per-plane spatial gradients, project them onto integrable surfaces via least squares, supervise them only in in-focus regions, and add a bidirectional soft monotonicity loss so focus probabilities peak at one plane and decay on both sides.
Key results
- On NYU Depth v2, RMSE 0.075 and AbsRel 0.013 versus DFV's 0.094 and 0.020 (about 20% and 35% reduction).
- On FoD500, Bump 1.31 versus DFV's 1.43, and best on all listed metrics.
- On DDFF 12-Scene, RMSE 0.0194 and MSE 4.7e-4 versus HybridDepth's 0.0200 and 5.1e-4; ablation shows removing L_fv raises RMSE from 0.075 to 0.078 on NYU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DualFocus instead of DFV | NYU Depth v2 · NYU Depth v2 | RMSE: 0.094 -> 0.075 | Helped |
| use DualFocus instead of DFV | FoD500 · FoD500 | Bump: 1.43 -> 1.31 | Helped |
| use DualFocus instead of HybridDepth | DDFF 12-Scene · DDFF 12-Scene | MSE: $5.1\times 10^{-4}$ -> $\mathbf{4.7} \bm{\times} \mathbf{10}^{\mathbf{-4}}$ | Helped |
| use DualFocus instead of HybridDepth | ARKitScenes validation · ARKitScenes validation | RMSE: 0.29 -> 0.28 | Helped |
| add spatial and focal variational constraints (ablation) | DualFocus · NYU Depth v2 | RMSE: 0.094 -> 0.075 | Helped |
| add spatial variational constraints (ablation) | DualFocus · NYU Depth v2 | RMSE: 0.090 -> 0.075 | Helped |
| add focal variational constraints (ablation) | DualFocus · NYU Depth v2 | RMSE: 0.078 -> 0.075 | Helped |
| project gradients onto integrable surfaces instead of directly supervising gradients (ablation) | DualFocus · NYU Depth v2 | RMSE: 0.083 -> 0.075 | Helped |
| use sharpness weight q rather than no weighting (ablation) | DualFocus spatial variational loss · NYU Depth v2 | RMSE: 0.077 -> 0.075 | Helped |
| use sharpness weight q rather than blurness weight 1-q (ablation) | DualFocus spatial variational loss · NYU Depth v2 | RMSE: 0.079 -> 0.075 | Helped |
| increase gradient/surface resolution from 10 × 10 to 14 × 14 (ablation) | DualFocus · NYU Depth v2 | RMSE: 0.079 -> 0.075 | Helped |
| increase gradient/surface resolution from 14 × 14 to 20 × 20 (ablation) | DualFocus · NYU Depth v2 | RMSE: 0.075 -> 0.073 | Helped |
Surprising findings
- Removing the focal variational loss increases the invalid focus trend rate from 6.2% to 46.2% on NYU, and from 13.4% to 88.2% on DDFF 12-Scene.
Limitations
- Weak depth cues in low-texture-contrast regions.
- Assumes spatially aligned focal stacks, sensitive to motion-induced misalignment.
- Blur characteristics specific to scenes or lenses may limit generalization.
Experimental setup
Datasets: NYU Depth v2 FoD500 DDFF 12-Scene ARKitScenes
Benchmarks: NYU Depth v2 FoD500 DDFF 12-Scene ARKitScenes
Metrics: MSE RMSE AbsRel SqRel delta1 delta2 delta3 Bumpiness log RMSE
Baselines: ZoeDepth VPD Marigold ECoDepth Depth Anything DefocusNet HybridDepth DFV RDF DDFF DistDepth ZeroDepth
Builds on: DFV VA-DepthNet HybridDepth
Closest papers by meaning
- Deep Depth From Focus With Differential Focus Volume (CVPR 2022 · similarity 0.90)
- MultiDiff: Consistent Novel View Synthesis from a Single Image (CVPR 2024 · similarity 0.84)
- Multi-Focus Image Fusion via Explicit Defocus Blur Modelling (AAAI 2025 · similarity 0.84)
- DepthMaster: Unified Monocular Depth Estimation for Perspective and Panoramic Images (NeurIPS 2026 · similarity 0.84)
- LDP: Language-driven Dual-Pixel Image Defocus Deblurring Network (CVPR 2024 · similarity 0.84)
- DepthFM: Fast Generative Monocular Depth Estimation with Flow Matching (AAAI 2025 · similarity 0.84)
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.