Research map / 3D vision / Depth and stereo
Blurry-Edges: Photon-Limited Depth Estimation from Defocused Boundaries
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 Blurry-Edges, a parametric wedge-based patch representation, and a two-stage CNN plus Transformer network that estimates depth from defocused boundaries in photon-limited image pairs via a closed-form DfD relation.
Problem
Depth from defocus is highly sensitive to image noise, so existing DfD methods fail on photon-limited, low-light images.
Approach
Model each image patch as alpha-clipped stacked wedges with per-wedge vertex, color, and boundary smoothness, then derive a closed-form depth equation from the smoothness of the same boundary in two differently defocused images; a local CNN and a global Transformer predict the representation.
Key results
- On the synthetic test set, sparse Ours achieves RMSE 5.281 cm and AbsRel 3.295 cm, versus 6.308 cm for Focal Track and 6.737 cm for Tang et al.
- Dense Ours-PP reaches delta1 0.806 and RMSE 3.992 cm, beating DefocusNet (6.092 cm) and DEReD (7.779 cm) on the synthetic set.
- The method handles about 4x higher noise standard deviation (18-19 LSB) than prior DfD methods, operating at 74-83 lux illuminance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Ours instead of Focal Track for sparse depth estimation | Depth from defocus · Synthetic testing set (sparse) | RMSE (cm): 6.308 -> 5.281 | Helped |
| Use Ours-PP instead of DefocusNet for dense depth estimation | Depth from defocus · Synthetic testing set (dense) | RMSE (cm): 6.092 -> 3.992 | Helped |
| Use patch size 11 × 11 instead of 21 × 21 (ablation) | Blurry-Edges depth estimation · Synthetic testing set | RMSE (cm): 5.281 -> 5.675 | Hurt |
| Use patch size 31 × 31 instead of 21 × 21 (ablation) | Blurry-Edges depth estimation · Synthetic testing set | RMSE (cm): 5.281 -> 6.123 | Hurt |
Surprising findings
- Analytical-style baselines retrained on the same noisy data still outperform some learning-based dense methods on sparse metrics, while the dense Ours-W variant is worse than Focal Track on RMSE.
Limitations
- Uses only two wedges (l=2) fixed for all experiments
- Depth is sparse along boundaries; dense maps require extra densification (Ours-W or U-Net post-processing)
- Real-world evaluation is limited to a single prototype camera with a deformable lens and a small number of scenes
Experimental setup
Datasets: Painting dataset MS-COCO
Benchmarks: synthetic Blurry-Edges test set (200 scenes from Painting and MS-COCO) real-world low-light captured image pairs
Metrics: delta1 delta2 delta3 RMSE AbsRel
Baselines: Focal Track Tang et al. PhaseCam3D DefocusNet DFV-DFF DEReD
Builds on: Guo et al. differential defocus depth estimation field-of-junction U-Net
Closest papers by meaning
- Multi-Focus Image Fusion via Explicit Defocus Blur Modelling (AAAI 2025 · similarity 0.87)
- Self-Distilled Depth Refinement with Noisy Poisson Fusion (NeurIPS 2024 · similarity 0.86)
- Residual Diffusion Deblurring Model for Single Image Defocus Deblurring (AAAI 2025 · similarity 0.86)
- Repurposing Marigold for Zero-Shot Metric Depth Estimation via Defocus Blur Cues (NeurIPS 2025 · similarity 0.85)
- $\alpha$Depth: Learning Single-Pass Soft Boundary Decomposition for Stereo Conversion (NeurIPS 2026 · similarity 0.85)
- Quad-Pixel Image Defocus Deblurring: A New Benchmark and Model (CVPR 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.