Research map / Image generation / Super-resolution
SGNet: Structure Guided Network via Gradient-Frequency Awareness for Depth Map 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 SGNet, a guided depth super-resolution network that adds gradient-domain calibration (GCM) and frequency-domain spectrum differencing (FAM/SDB) to transfer RGB high-frequency structure into LR depth.
Problem
Guided depth map super-resolution (DSR) recovers blurry, low-frequency-dominated LR depth structure; prior RGB-guided methods operate mainly in the spatial domain.
Approach
Exploit gradient and frequency (DFT amplitude/phase) domains: a gradient calibration module sharpens LR depth using RGB gradient priors, and recursive spectrum differencing blocks propagate RGB high-frequency components into depth, with gradient- and frequency-aware losses.
Key results
- On NYU-v2 x16, SGNet reaches 4.77 RMSE vs. 4.86 for the second-best SUFT.
- On RGB-D-D x16, SGNet reaches 2.55 RMSE vs. 2.71 for SUFT.
- On real-world RGB-D-D retrained setting, SGNet* achieves 5.32 RMSE vs. 5.41 for SUFT*.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SGNet instead of the strongest baseline | Depth super-resolution · NYU-v2 ×4 | RMSE (cm): 1.12 -> 1.10 | Helped |
| Use SGNet instead of the strongest baseline | Depth super-resolution · NYU-v2 ×8 | RMSE (cm): 2.51 -> 2.44 | Helped |
| Use SGNet instead of the strongest baseline | Depth super-resolution · NYU-v2 ×16 | RMSE (cm): 4.86 -> 4.77 | Helped |
| Use SGNet instead of the strongest baseline | Depth super-resolution · RGB-D-D ×4 | RMSE (cm): 1.08 -> 1.10 | Hurt |
| Use SGNet instead of the strongest baseline | Depth super-resolution · RGB-D-D ×8 | RMSE (cm): 1.69 -> 1.64 | Helped |
| Use SGNet instead of the strongest baseline | Depth super-resolution · RGB-D-D ×16 | RMSE (cm): 2.71 -> 2.55 | Helped |
| Use SGNet instead of the strongest baseline | Depth super-resolution · Middlebury ×4 | RMSE (cm): 1.07 -> 1.15 | Hurt |
| Use SGNet instead of the strongest baseline | Depth super-resolution · Middlebury ×8 | RMSE (cm): 1.74 -> 1.64 | Helped |
| Use SGNet instead of the strongest baseline | Depth super-resolution · Middlebury ×16 | RMSE (cm): 3.18 -> 2.95 | Helped |
| Use SGNet instead of the strongest baseline | Depth super-resolution · Lu ×4 | RMSE (cm): 0.79 -> 1.03 | Hurt |
| Use SGNet instead of the strongest baseline | Depth super-resolution · Lu ×8 | RMSE (cm): 1.67 -> 1.61 | Helped |
| Use SGNet instead of the strongest baseline | Depth super-resolution · Lu ×16 | RMSE (cm): 3.92 -> 3.55 | Helped |
Limitations
- Gains on Middlebury x4 and Lu x4 are not best (SUFT and RSAG are better there).
- Added complexity relative to baseline (FLOPs rise by about 6%).
Experimental setup
Datasets: NYU-v2 RGB-D-D Middlebury Lu
Benchmarks: NYU-v2 Middlebury Lu RGB-D-D
Metrics: RMSE (cm)
Baselines: TGV SDF DJF DJFR PAC DMSG GbFT DKN FDKN FDSR JIIF CTKT
Builds on: DKN FDSR DCTNet RSAG SUFT
Closest papers by meaning
- Structure Flow-Guided Network for Real Depth Super-resolution (AAAI 2023 · similarity 0.91)
- Recurrent Structure Attention Guidance for Depth Super-resolution (AAAI 2023 · similarity 0.91)
- Intrinsic Phase-Preserving Networks for Depth Super Resolution (AAAI 2024 · similarity 0.89)
- DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution (CVPR 2025 · similarity 0.86)
- DegBins: Degradation-Driven Binning for Depth Super-Resolution (NeurIPS 2026 · similarity 0.86)
- Completion as Enhancement: A Degradation-Aware Selective Image Guided Network for Depth Completion (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.