Research map / 3D vision / Depth and stereo
Decoupling Fine Detail and Global Geometry for Compressed 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 GDNet, which decouples compressed depth map super-resolution into fine-detail features from the RGB image (FGDE) and global geometry from the compressed depth via a low-rank projection encoder (GGE), trained with SILog loss.
Problem
Recovering high-quality depth maps from compressed sources (downsampled, bit-depth compressed, and noisy), where bit-depth compression flattens subtle depth variations and random noise corrupts global geometry estimation.
Approach
Decouple reconstruction into two streams: a fine geometry detail encoder that aggregates high-resolution low-level RGB features enriched with low-resolution context features via self- and cross-attention, and a global geometry encoder that projects compressed-depth features into a low-rank space (via QR decomposition and a Neumann-series approximated projection) to suppress noise.
Key results
- On the AIM 2024 CDU dataset, GDNet reaches MAE 0.212 and RMSE 0.375, versus 0.294/0.432 for the next best DAS-Depth.
- On Compressed-NYU, GDNet achieves MAE 0.0322 and RMSE 0.0739, versus 0.0426/0.1051 for SGNet.
- Ablation on Compressed-NYU shows removing FGDE or GGE increases error, and the SILog loss outperforms L1 and MSE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GDNet instead of the strongest listed baseline, DAS-Depth | Compressed depth map super-resolution · AIM 2024 Compressed Depth Upsampling Challenge | MAE: 0.294 -> 0.212 | Helped |
| Use GDNet instead of the strongest listed baseline, DAS-Depth | Compressed depth map super-resolution · AIM 2024 Compressed Depth Upsampling Challenge | RMSE: 0.432 -> 0.375 | Helped |
| Use GDNet instead of the strongest listed baseline, SGNet | Compressed depth map super-resolution · Compressed-NYU | MAE: 0.0426 -> 0.0322 | Helped |
| Use GDNet instead of the strongest listed baseline, SGNet | Compressed depth map super-resolution · Compressed-NYU | RMSE: 0.1051 -> 0.0739 | Helped |
| Add FGDE (ablation) | GDNet · Compressed-NYU | MAE | Helped |
| Add GGE (ablation) | GDNet · Compressed-NYU | MAE | Helped |
| Use SILog loss instead of L1 loss (ablation) | GDNet · Compressed-NYU | MAE: 0.0347 -> 0.0322 | Helped |
| Use SILog loss instead of MSE loss (ablation) | GDNet · Compressed-NYU | MAE: 0.0372 -> 0.0322 | Helped |
| Use 1 self-attention and 1 cross-attention module (ablation) | GDNet FGDE · Compressed-NYU | MAE: 0.0357 -> 0.0322 | Helped |
| Use 1 self-attention and 1 cross-attention module (ablation) | GDNet FGDE · Compressed-NYU | MAE: 0.0385 -> 0.0322 | Helped |
| Use 1 self-attention and 1 cross-attention module (ablation) | GDNet FGDE · Compressed-NYU | MAE: 0.0345 -> 0.0322 | Helped |
| Use 1 self-attention and 1 cross-attention module (ablation) | GDNet FGDE · Compressed-NYU | MAE: 0.0372 -> 0.0322 | Helped |
Limitations
- Evaluated only on two datasets, one synthetic (Compressed-NYU)
- Reported RMSE for Compressed-NYU in text (0.739) is inconsistent with table (0.0739)
Experimental setup
Datasets: AIM 2024 Compressed Depth Upsampling Challenge dataset (from TartanAir) Compressed-NYU (synthesized from NYU Depth V2) NYU Depth Dataset V2 TartanAir
Benchmarks: AIM 2024 Compressed Depth Upsampling Challenge dataset Compressed-NYU
Metrics: MAE RMSE
Baselines: DAS-Depth DINOv2-ControlNet RAFT-DU DAv2 ++ SGNet Depth Anything Depth Anything V2 UniDepth iDisc AdaBins
Builds on: Depth Anything V2 SGNet AdaBins
Closest papers by meaning
- Improving Depth Completion via Depth Feature Upsampling (CVPR 2024 · similarity 0.87)
- Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (ICLR 2025 · similarity 0.85)
- DegBins: Degradation-Driven Binning for Depth Super-Resolution (NeurIPS 2026 · similarity 0.85)
- GeoDepth: From Point-to-Depth to Plane-to-Depth Modeling for Self-Supervised Monocular Depth Estimation (CVPR 2025 · similarity 0.84)
- SGNet: Structure Guided Network via Gradient-Frequency Awareness for Depth Map Super-resolution (AAAI 2024 · similarity 0.84)
- D$^2$GS: Dense Depth Regularization for LiDAR-free Urban Scene Reconstruction (NeurIPS 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.