Research map / Label-efficient and robust learning / Noisy labels
Self-Distilled Depth Refinement with Noisy Poisson Fusion
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.
Models depth refinement as a noisy Poisson fusion problem and proposes SDDR, which uses coarse-to-fine self-distilled depth edge pseudo-labels and edge-based losses to get a one-stage or efficient two-stage refiner.
Problem
Depth refinement to produce high-resolution depth with sharp edges and consistent structure, where tile-based two-stage methods are costly and inconsistent, and prior training labels are noisy and blurred, limiting generalization.
Approach
Decouples depth prediction errors into local inconsistency and edge deformation noise, then uses coarse-to-fine self-distillation to produce low-noise depth edge pseudo-labels, supervised via an edge-guided gradient loss and a quantile-based edge-based fusion loss that learns a soft fusion mask.
Key results
- On Middlebury2021 with ZoeDepth, SDDR reaches delta1 0.913 and REL 0.096 versus Boost's 0.911 and 0.099 (two-stage comparison).
- With MiDaS on Middlebury2021, REL improves 6.6% and ORD 20.7% over Kim et al. in the one-stage setting.
- Ablation on Middlebury2021 with LeReS: adding L_grad and L_fusion improves D3R by 10.0% and REL by 3.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SDDR instead of Boost with ZoeDepth | Depth refinement · Hypersim | delta_1: 0.881 -> 0.933 | Helped |
| Use SDDR instead of Boost with ZoeDepth | Depth refinement · Middlebury2021 | ORD: 0.210 -> 0.202 | Helped |
| Increase coarse-to-fine refinement from S=0 to S=3 (ablation) | SDDR with LeReS · Middlebury2021 | ORD: 0.313 -> 0.305 | Helped |
| Add edge-guided gradient loss to ground-truth supervision (ablation) | SDDR with LeReS · Middlebury2021 | ORD: 0.313 -> 0.307 | Helped |
| Add edge-based fusion loss to ground-truth supervision (ablation) | SDDR with LeReS · Middlebury2021 | REL: 0.124 -> 0.120 | Helped |
| Use both edge-guided gradient loss and edge-based fusion loss (ablation) | SDDR with LeReS · Middlebury2021 | ORD: 0.313 -> 0.305 | Helped |
| Train GBDF with SDDR-produced edge representation G_S (ablation) | GBDF with LeReS · Middlebury2021 | D3R: 0.258 -> 0.230 | Helped |
Surprising findings
- Training another model (GBDF) with the self-distilled edge labels G_S improves its accuracy and edge quality, suggesting the labels transfer to other refiners.
Limitations
- Performance degrades on inputs with specular surfaces, low light, or weak textures where the depth predictor is poor.
- Refinement network architecture could be explored further.
Experimental setup
Datasets: TartanAir Irs UnrealStereo4K MVS-Synth Holopix50K iBims-1 WSVD VDW HRWSI
Benchmarks: Middlebury2021 Multiscopic Hypersim DIML DIODE
Metrics: delta1 REL ORD D3R Abs Rel RMSE FLOPs inference time
Baselines: MiDaS LeReS ZoeDepth Kim et al. Graph-GDSR GBDF Boost PatchFusion
Builds on: Boost GBDF MiDaS LeReS ZoeDepth
Closest papers by meaning
- Blurry-Edges: Photon-Limited Depth Estimation from Defocused Boundaries (CVPR 2025 · similarity 0.86)
- Mind The Edge: Refining Depth Edges in Sparsely-Supervised Monocular Depth Estimation (CVPR 2024 · similarity 0.86)
- Distilling Monocular Foundation Model for Fine-grained Depth Completion (CVPR 2025 · similarity 0.85)
- $\alpha$Depth: Learning Single-Pass Soft Boundary Decomposition for Stereo Conversion (NeurIPS 2026 · similarity 0.85)
- PatchFusion: An End-to-End Tile-Based Framework for High-Resolution Monocular Metric Depth Estimation (CVPR 2024 · similarity 0.85)
- Pixel-Perfect Depth with Semantics-Prompted Diffusion Transformers (NeurIPS 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.