Research map / 3D vision / Depth and stereo
Completion as Enhancement: A Degradation-Aware Selective Image Guided Network for Depth Completion
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 SigNet, which reframes depth completion as depth enhancement by densifying sparse depth with non-CNN tools and fusing RGB via degradation-aware decomposition and a conditional Mamba.
Problem
Depth completion from sparse, irregularly sampled depth suffers from mismatch and ambiguity when convolutions are applied directly to incomplete depth data.
Approach
Densify sparse depth with non-CNN tools (IP-Basic, fast marching) to get coarse dense depth, then model an implicit self-supervised degradation between coarse and target depth; decompose the degradation with DCT into patches that select RGB high-frequency features and condition a Mamba state-space fusion.
Key results
- On DIML, SigNet reaches RMSE 55 mm and REL 0.006, versus 68 mm RMSE for the second-best BPNet (about 19% RMSE reduction).
- On NYUv2, SigNet reports RMSE 83 mm with 3.3M parameters versus 88 mm for BPNet with 89.9M parameters, a 96% parameter reduction.
- On TOFDC, SigNet reduces RMSE by 5 mm relative to TPVD while using 27.9M fewer parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SigNet instead of BPNet | depth completion · NYUv2 | RMSE (mm): 88 -> 83 | Helped |
| use SigNet instead of AGG-Net | depth completion · DIML | RMSE (mm): 78 -> 55 | Helped |
| use SigNet instead of BPNet | depth completion · SUN RGBD | RMSE (mm): 124 -> 116 | Helped |
| use SigNet instead of TPVD | depth completion · TOFDC | RMSE (mm): 92 -> 87 | Helped |
| add densification and a degradation bridge, and concatenate RGB and degradation features (ablation) | SigNet-i on DIML · DIML test split | RMSE (mm): 108 -> 95 | Helped |
| add Gaussian denoising (ablation) | SigNet-ii on DIML · DIML test split | RMSE (mm): 95 -> 92 | Helped |
| add sigmoid attention and element-wise RGB feature selection (ablation) | SigNet-iii on DIML · DIML test split | RMSE (mm): 92 -> 74 | Helped |
| add degradation-aware decomposition (ablation) | SigNet-iv on DIML · DIML test split | RMSE (mm): 74 -> 68 | Helped |
| add conditional Mamba fusion (ablation) | SigNet-v on DIML · DIML test split | RMSE (mm): 68 -> 55 | Helped |
| replace vanilla Vision Mamba with conditional Mamba conditioned on degradation and RGB-D features (ablation) | DADF on DIML · DIML | RMSE (mm): 8 mm | Helped |
| add DCT-based degradation decomposition (ablation) | conditional Mamba fusion on DIML · DIML | RMSE (mm) | Helped |
| increase DADF iterations (ablation) | DADF on DIML · DIML | RMSE (mm) | Helped |
Surprising findings
- Degradation estimation fails on the outdoor KITTI benchmark because sparse ground truth and long distances weaken edge constraints.
Limitations
- Poor degradation estimation on outdoor KITTI with sparse ground truth and long depth ranges
- Evaluated mainly on indoor and simple outdoor scenes
Experimental setup
Datasets: NYUv2 DIML SUN RGBD TOFDC
Benchmarks: NYUv2 DIML SUN RGBD TOFDC
Metrics: RMSE REL delta1 delta2 delta3 parameters inference time
Baselines: Bilateral filtering S2D CSPN DfuseNet DM-LRN RDF-GAN AGG-Net BPNet TPVD GuideNet PENet NLSPN
Builds on: Vision Mamba degradation learning for blind super-resolution AGG-Net IP-Basic
Closest papers by meaning
- Improving Depth Completion via Depth Feature Upsampling (CVPR 2024 · similarity 0.86)
- Zero-shot Depth Completion via Test-time Alignment with Affine-invariant Depth Prior (AAAI 2025 · similarity 0.86)
- DeCoTR: Enhancing Depth Completion with 2D and 3D Attentions (CVPR 2024 · similarity 0.86)
- Efficient Indoor Depth Completion Network Using Mask-adaptive Gated Convolution (AAAI 2025 · similarity 0.86)
- Bilateral Propagation Network for Depth Completion (CVPR 2024 · similarity 0.86)
- DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution (CVPR 2025 · 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.