Research map / Image generation / Super-resolution
DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution
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 DORNet, a blind RGB-guided depth super-resolution network that learns implicit, label-free degradation representations via routing-based degradation regularization and fuses RGB via degradation-oriented feature transformation.
Problem
Depth super-resolution under unknown, real-world depth degradation (sensor limits, low-reflectance surfaces, illumination), where methods assuming fixed bicubic degradation degrade sharply.
Approach
Self-supervised degradation learning: a router selects multi-scale degradation kernel generators that re-degrade the predicted HR depth to match the LR depth (plus contrastive loss), and the learned degradation representation steers deformable-conv RGB feature selection (DOFT) for RGB-D fusion.
Key results
- On real-world RGB-D-D, DORNet reaches 3.68 cm RMSE versus 4.08 cm for SFG (noise-free setting in Table 2 variant), with 3.05M parameters.
- On TOFDSR, DORNet lowers RMSE by 0.12 cm relative to SGNet (4.33 to 4.21 cm in Table 1).
- On NYU-v2 x4, DORNet-T (0.46M) beats DCTNet by 16% RMSE; DORNet uses 91% fewer parameters than SGNet with 8% higher RMSE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DORNet instead of SFG | Real-world depth super-resolution · RGB-D-D | RMSE (cm): 3.88 -> 3.42 | Helped |
| Use DORNet instead of SGNet | Real-world depth super-resolution · TOFDSR | RMSE (cm): 4.33 -> 4.21 | Helped |
| Use DORNet instead of SFG | Noisy depth super-resolution · RGB-D-D, w/ Noisy | RMSE (cm): 4.08 -> 3.68 | Helped |
| Use DORNet instead of DCTNet | Noisy depth super-resolution · TOFDSR, w/ Noisy | RMSE (cm): 5.46 -> 4.47 | Helped |
| Use DORNet instead of the strongest baseline | Synthetic depth super-resolution · NYU-v2 ×4, ×8, and ×16 | RMSE | Hurt |
| Add degradation learning (DL) to the baseline (ablation) | DORNet · RGB-D-D | RMSE | Helped |
| Add degradation regularization (DR) to degradation learning (ablation) | DORNet · RGB-D-D | RMSE: 0.82cm (w/o Noisy); 0.83cm (w/ Noisy) | Helped |
| Replace DORNet degradation learning and regularization with DASR (ablation) | DORNet · RGB-D-D, w/ Noisy | RMSE (cm): 3.69 -> 4.06 | Hurt |
| Replace DORNet degradation learning and regularization with KDSR (ablation) | DORNet · RGB-D-D, w/ Noisy | RMSE (cm): 3.69 -> 3.86 | Hurt |
| Add degradation loss Ldeg (ablation) | DORNet with reconstruction loss · RGB-D-D | RMSE | Helped |
| Add contrastive loss Lcont (ablation) | DORNet with reconstruction loss · RGB-D-D | RMSE | Helped |
| Add both Ldeg and Lcont (ablation) | DORNet with reconstruction loss · RGB-D-D | RMSE: 0.20cm (w/o Noisy); 0.27cm (w/ Noisy) | Helped |
Surprising findings
- On synthetic NYU-v2 x16, DORNet is slightly worse than SGNet and several others in RMSE (5.60 vs 4.77), indicating limited gains at large scales.
Limitations
- Gains on synthetic NYU-v2 at large upsampling factors (x16) are not consistent with the best baselines.
- Degradation learning is evaluated mainly on a single real dataset captured with one ToF sensor (Huawei P30 Pro).
Experimental setup
Datasets: RGB-D-D TOFDSR NYU-v2
Benchmarks: RGB-D-D TOFDSR NYU-v2
Metrics: RMSE (cm) parameters (M)
Baselines: DJF DJFR CUNet DKN FDKN FDSR DCTNet SUFT SSDNet SFG SGNet GraphSR
Builds on: DKN FDSR DASR KDSR VGG19
Closest papers by meaning
- DegBins: Degradation-Driven Binning for Depth Super-Resolution (NeurIPS 2026 · similarity 0.90)
- Unsupervised Degradation Representation Aware Transform for Real-World Blind Image Super-Resolution (AAAI 2025 · similarity 0.89)
- SGNet: Structure Guided Network via Gradient-Frequency Awareness for Depth Map Super-resolution (AAAI 2024 · similarity 0.86)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.86)
- Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning (CVPR 2024 · similarity 0.86)
- Navigating Beyond Dropout: An Intriguing Solution towards Generalizable Image Super Resolution (CVPR 2024 · similarity 0.86)
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.