Research map / Image generation / Super-resolution
DegBins: Degradation-Driven Binning 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 DegBins, a multi-stage hybrid classification-regression depth super-resolution framework whose bin ranges and probabilities are adapted using high-dimensional degradation representations from DORNet.
Problem
Depth super-resolution with color guidance is usually posed as additive residual regression in a low-dimensional feature space, which poorly captures complex, spatially varying HR-LR degradation relationships.
Approach
Reformulates DSR as hybrid classification-regression over pixel-wise adaptive depth bins whose ranges and probabilities are modulated by degradation representations (variance-based range widening, DCN-modulated GRU probability estimation), refined over four coarse-to-fine stages.
Key results
- On synthetic NYU v2 x16, DegBins reaches RMSE 7.11 vs 7.37 for the DuCos-based baseline (ablation i) and second-best DuCos at 7.37 in the main table.
- On real-world RGB-D-D, DegBins achieves RMSE 3.52 cm, MAE 1.45, delta1 98.01, improving over DuCos (3.68).
- On compressed DSR NYU v2, DegBins RMSE 51.26 vs 58.74 for DuCos; a lightweight residual-group variant reduces parameters from 42.15M to 16.28M with RMSE 3.63 on RGB-D-D.
Surprising findings
- A lightweight residual-group backbone variant (DegBins-RG) attains lower RMSE (3.63) than the DuCos-based DuCos baseline (3.68) on RGB-D-D while using far fewer parameters.
Limitations
- High computational complexity due to heavy DuCos/Depth Anything v2 feature backbone; limited real-time performance.
- Degradation representation relies on a pretrained DORNet block, adding dependency on a separate pretrained component.
- Gains over strongest baseline (DuCos) are often small (e.g., 0.08 cm RMSE on NYU v2 x2).
Experimental setup
Datasets: Hypersim Middlebury Lu NYU v2 RGB-D-D TOFDSR
Benchmarks: Middlebury Lu NYU v2 RGB-D-D TOFDSR
Metrics: RMSE MAE delta_1
Baselines: DJF DJFR CUNet DKN FDKN FDSR DCTNet DADA DORNet DuCos SUFT SFG
Builds on: AdaBins DORNet DuCos IEBins LocalBins BinsFormer
Closest papers by meaning
- DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution (CVPR 2025 · similarity 0.90)
- Unsupervised Degradation Representation Aware Transform for Real-World Blind Image Super-Resolution (AAAI 2025 · similarity 0.86)
- Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution (AAAI 2025 · similarity 0.86)
- Intrinsic Phase-Preserving Networks for Depth Super Resolution (AAAI 2024 · similarity 0.86)
- SGNet: Structure Guided Network via Gradient-Frequency Awareness for Depth Map Super-resolution (AAAI 2024 · similarity 0.86)
- Improving Depth Completion via Depth Feature Upsampling (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.