Research map / Image generation / Super-resolution
C2PD: Continuity-Constrained Pixelwise Deformation for Guided 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 C2PD, which treats depth maps as ideal plastic material and deforms upsampled low-resolution depth guided by RGB using continuity-constrained pixelwise operations (CAPO) and gradient-based PCGD for guided depth super-resolution.
Problem
Guided depth super-resolution (GDSR) methods treat depth maps as images with discretely computed values, which fails to restore the continuity inherent in depth maps, especially at large scale factors.
Approach
Model depth as an ideal plastic object: CAPO is a cross-modal operation that redistributes values under volume-conserving constraints guided by RGB-derived information, and PCGD applies CAPO to depth gradients in x and y directions and integrates them, so edges are moved rather than blended.
Key results
- Achieves best RMSE on Middlebury x8 (1.57), x16 (2.80), Lu x4-x16, NYU_v2 x8 (2.36) and x16 (4.48) among compared methods.
- At x32 scale, reports RMSE 5.68 (Middlebury), 6.48 (Lu), 9.03 (NYU_v2), 4.11 (RGBDD) versus DADA's 6.94, 8.47, 10.81, 4.96.
- Generalization on real-world RGBDD (x32, NYU_v2-trained): RMSE 6.68 vs SUFT/SGNet 7.22.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use C2PD instead of the strongest listed baseline | C2PD · Middlebury x16 | RMSE (cm): 2.95 -> 2.80 | Helped |
| Use C2PD instead of the strongest listed baseline | C2PD · Lu x16 | RMSE (cm): 3.55 -> 3.11 | Helped |
| Use C2PD instead of the strongest listed baseline | C2PD · NYU_v2 x4 | RMSE (cm): 1.10 -> 1.25 | Hurt |
| Use C2PD instead of the strongest listed baseline | C2PD · RGBDD x16 | RMSE (cm): 2.55 -> 2.41 | Helped |
| Use C2PD instead of DADA | C2PD · NYU_v2 x32 | RMSE (cm): 10.81 -> 9.03 | Helped |
| Use C2PD instead of SUFT or SGNet | C2PD · Real-world RGBDD | RMSE: 7.22 -> 6.68 | Helped |
| Replace CAPO with residual convolution (ablation) | C2PD · GDSR; Fig. 7(a) | performance | Hurt |
| Set CAPO length to 4 rather than other tested lengths (ablation) | C2PD PCGD · GDSR; Fig. 7(b) | performance | Helped |
| Operate on both gradient directions in parallel with shared parameters (ablation) | C2PD PCGD · x8 GDSR; Fig. 8(a) | performance | Helped |
Surprising findings
- At x4 the method is often suboptimal relative to baselines, which the authors attribute to not using many parameters for dense feature reconstruction.
Limitations
- Performance at x4 is not best across datasets, as the method uses fewer parameters for dense reconstruction.
- Mathematical motivation relies on loosely justified analogies to plastic deformation and continuity; discretization details are informal.
- Ablation on CAPO kernel length used dot products for lengths other than 4, potentially confounding comparisons.
Experimental setup
Datasets: NYU_v2
Benchmarks: Middlebury Lu NYU_v2 RGBDD
Metrics: RMSE (cm)
Baselines: DKN FDKN FDSR JIIF DCTNet SUFT SSDNet DAGF DADA SGNet
Builds on: DADA U-Net DKN FDSR
Closest papers by meaning
- SGNet: Structure Guided Network via Gradient-Frequency Awareness for Depth Map Super-resolution (AAAI 2024 · similarity 0.84)
- Leveraging Global Stereo Consistency for Category-Level Shape and 6D Pose Estimation from Stereo Images (CVPR 2025 · similarity 0.84)
- SD-MVS: Segmentation-Driven Deformation Multi-View Stereo with Spherical Refinement and EM Optimization (AAAI 2024 · similarity 0.84)
- Decoupling Fine Detail and Global Geometry for Compressed Depth Map Super-Resolution (CVPR 2025 · similarity 0.83)
- DVP-MVS: Synergize Depth-Edge and Visibility Prior for Multi-View Stereo (AAAI 2025 · similarity 0.83)
- Intrinsic Phase-Preserving Networks for Depth Super Resolution (AAAI 2024 · similarity 0.82)
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.