Research map / 3D vision / Depth and stereo
A Simple yet Universal Framework 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.
Defines Universal Depth Completion (UniDC) and proposes a hyperbolic, foundation-model-based few-shot depth completion framework that adapts a MiDaS backbone to arbitrary depth sensors using minimal labeled data.
Problem
Depth completion generalizes poorly across scenes and sensors with different scanning ranges and densities, and dense metric ground truth is expensive, so models need to adapt from few labeled examples.
Approach
Use a monocular depth foundation model (MiDaS) as a frozen-ish feature extractor with bias tuning instead of learning a new depth encoder, perform sparse-to-dense conversion with a bilateral-style hyperbolic kernel, and refine with CSPN++-style propagation using multi-curvature hyperbolic affinity maps whose curvature is predicted per scene.
Key results
- On NYUv2 1-shot, the method reaches RMSE 0.2099 vs 0.3573 for BPNet and 0.3583 for DepthPrompting.
- On KITTI DC 1-shot, RMSE 1.6840 vs 2.9561 for DepthPrompting (the strongest reported baseline there).
- Hyperbolic embedding gives zero-shot NYU RMSE 0.323 vs 13.889 for Euclidean, in the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UniDC with the foundation model and proposed completion framework instead of the strongest listed baseline | UniDC · NYUv2 1-shot | RMSE: 0.3573 -> 0.2099 | Helped |
| use UniDC with the foundation model and proposed completion framework instead of DepthPrompting | UniDC · NYUv2 10-shot | RMSE: 0.2195 -> 0.1657 | Helped |
| use UniDC with the foundation model and proposed completion framework instead of the strongest listed baseline | UniDC · NYUv2 100-shot | RMSE: 0.1757 -> 0.1473 | Helped |
| use UniDC with the foundation model and proposed completion framework instead of the strongest listed baseline | UniDC · KITTI DC 1-shot | RMSE: 2.6094 -> 1.6840 | Helped |
| use UniDC with the foundation model and proposed completion framework instead of the strongest listed baseline | UniDC · KITTI DC 10-shot | RMSE: 1.8965 -> 1.3850 | Helped |
| use UniDC with the foundation model and proposed completion framework instead of the strongest listed baseline | UniDC · KITTI DC 100-shot | RMSE: 1.3126 -> 1.2238 | Helped |
| add feature fusion (Eq.3) (ablation) | UniDC · KITTI DC 1-shot | RMSE: 2.17±0.99 -> 1.68±0.07 | Helped |
| add hyperbolic embedding (Eq.5) (ablation) | UniDC · KITTI DC 1-shot | RMSE: 1.85±0.32 -> 1.68±0.07 | Helped |
| add curvature generation (Eq.6) (ablation) | UniDC · KITTI DC 1-shot | RMSE: 1.84±0.37 -> 1.68±0.07 | Helped |
| add scale-invariant loss (Eq.2) (ablation) | UniDC · KITTI DC 1-shot | RMSE: 2.31±1.66 -> 1.68±0.07 | Helped |
| replace Euclidean operations with hyperbolic operations (ablation) | depth completion · NYU w/o training | RMSE: 13.889 -> 0.323 | Helped |
| replace Euclidean operations with hyperbolic operations (ablation) | depth completion · KITTI w/o training | RMSE: 77.642 -> 4.061 | Helped |
Surprising findings
- Directly feeding relative-scale Depth Foundation Model output into prior SPN frameworks (the 'VFM' variant) often underperforms the baseline, which the authors attribute to linear fitting issues.
- Convolutional MiDaS backbone outperformed ViT-based Depth Anything and UniDepth in 10- and 100-shot settings.
Limitations
- Only handles a single image plus one sparse depth map; no multi-view or full 3D reconstruction input.
- Not directly applicable to noisy, highly sparse radar depth without uncertainty estimation.
- Gains on the in-domain KITTI benchmark with full data are smaller; full-dataset results lag some methods (e.g., CompletionFormer Small on NYU and KITTI in Table 11).
Experimental setup
Datasets: NYUv2 KITTI DC SUN RGB-D
Benchmarks: NYUv2 KITTI DC SUN RGB-D
Metrics: RMSE MAE DELTA1
Baselines: CSPN NLSPN DySPN CompletionFormer CostDCNet BPNet DepthPrompting S2D LRRU DFU OGNI-DC
Builds on: CSPN++ MiDaS DepthPrompting Poincaré ball model bias tuning
Closest papers by meaning
- Zero-shot Depth Completion via Test-time Alignment with Affine-invariant Depth Prior (AAAI 2025 · similarity 0.90)
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.90)
- Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (ICLR 2025 · similarity 0.90)
- Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data (CVPR 2024 · similarity 0.90)
- Distilling Monocular Foundation Model for Fine-grained Depth Completion (CVPR 2025 · similarity 0.89)
- UniDepth: Universal Monocular Metric Depth Estimation (CVPR 2024 · similarity 0.89)
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.