Research map / Image generation / Super-resolution
Intrinsic Phase-Preserving Networks for 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 IPPNet, a guided depth super-resolution network with a Phase-Preserving Filtering Module that uses depth-derived filters to suppress RGB texture components lying outside the depth map's intrinsic phase.
Problem
Guided depth map super-resolution (DSR) is hurt by over-texture and halo artifacts when RGB edges/textures are naively fused into depth features.
Approach
Spectral analysis shows LR and HR depth maps share the same intrinsic phase while RGB contains extra phases; a phase-preserving filter generated from LR depth features prunes RGB components before depth is enhanced by modulation, in a one-directional depth-to-RGB-then-RGB-to-depth scheme.
Key results
- On NYU v2 16x, IPPNet achieves RMSE 4.69 vs. 5.27 for JIIF and 4.98 for PMBANet (best on 16x).
- On Lu 16x, IPPNet reaches RMSE 2.88, and on Middlebury 8x RMSE 1.51, both better than prior SOTA.
- On RGB-D-D real-world branch trained directly, IPPNet obtains RMSE 5.32 vs. 5.43 for DCTNet*, and it outperforms baselines in the NYU-trained no-retraining setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Remove RGB guidance (ablation) | IPPNet · NYU v2 | RMSE: 1.20 -> 1.92 (×4); 2.45 -> 3.91 (×8); 4.69 -> 7.62 (×16) | Hurt |
| Remove the filtering stream from PPFM (ablation) | IPPNet · NYU v2 | RMSE: 1.20 -> 1.37 (×4); 2.45 -> 2.84 (×8); 4.69 -> 5.61 (×16) | Hurt |
| Replace PPFM with simultaneous RGB-depth fusion (ablation) | IPPNet · NYU v2 | RMSE: 1.20 -> 1.32 (×4); 2.45 -> 2.76 (×8); 4.69 -> 5.18 (×16) | Hurt |
| Remove the refinement module (ablation) | IPPNet · NYU v2 | RMSE: 1.20 -> 1.36 (×4); 2.45 -> 2.82 (×8); 4.69 -> 5.45 (×16) | Hurt |
| Place PPFM after the refinement module (ablation) | IPPNet · NYU v2 | RMSE: 1.20 -> 1.29 (×4); 2.45 -> 2.64 (×8); 4.69 -> 4.92 (×16) | Hurt |
| Use modulation instead of addition, concatenation, deformable convolution, or SFT for cross-modality fusion (ablation) | IPPNet · NYU v2 8× | RMSE | Helped |
| Use three PPFM blocks (ablation) | IPPNet · NYU v2 8× | RMSE | No clear effect |
Surprising findings
- Filtering RGB before depth modulation performs better than fusing RGB and depth simultaneously, and placing PPFM after the refinement module degrades performance.
- Some baselines (e.g., PMBANet) are better than IPPNet at 4x on NYU v2 (1.06 vs 1.20), so IPPNet does not win every setting.
Limitations
- Phase-preserving motivation is supported mainly by spectral analyses on specific datasets; theoretical justification is heuristic (DFT analysis under assumed gain factor).
- Does not win all settings; PMBANet is better at 4x NYU v2 and some RGB-D-D settings.
- Real-world generalization evaluated only on RGB-D-D with a single real-world branch.
Experimental setup
Datasets: NYU v2 RGB-D-D Middlebury Lu Noisy Middlebury
Benchmarks: NYU v2 Middlebury Lu RGB-D-D Noisy Middlebury
Metrics: RMSE edge errors inference runtime parameters
Baselines: DKN FDKN JIIF FDSR DCTNet AHMF PMBANet PAC DJF
Builds on: DKN JIIF EDSR FDSR
Closest papers by meaning
- SGNet: Structure Guided Network via Gradient-Frequency Awareness for Depth Map Super-resolution (AAAI 2024 · similarity 0.89)
- Structure Flow-Guided Network for Real Depth Super-resolution (AAAI 2023 · similarity 0.89)
- Channel Attention Based Iterative Residual Learning for Depth Map Super-Resolution (CVPR 2020 · similarity 0.87)
- DegBins: Degradation-Driven Binning for Depth Super-Resolution (NeurIPS 2026 · similarity 0.86)
- DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution (CVPR 2025 · similarity 0.85)
- Completion as Enhancement: A Degradation-Aware Selective Image Guided Network for Depth Completion (CVPR 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.