Research map / 3D vision / Depth and stereo

Metric Depth Estimation from Arbitrarily Degraded Low-Resolution Depth Prompts

NeurIPS 2026 · 3D vision › Depth and stereo · approach: Self-supervised monocular depth

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 AdaDS, a two-stage diffusion-based depth super-resolution framework that uses uncertainty-aware calibration to adaptively select a diffusion start timestep and injects learned noise before denoising with a pre-trained Marigold-LCM.

Problem

Depth super-resolution from arbitrarily degraded low-resolution depth prompts with RGB guidance; existing regression-based methods generalize poorly under severe or unknown degradations and arbitrary upsampling factors.

Approach

Exploits the contraction of Gaussian forward diffusion: a calibration stage predicts a refined depth latent and per-pixel uncertainty, which sets an adaptive timestep; a sampling module then injects timestep-conditioned noise so the sample lies in the high-density region of the depth posterior, and a pre-trained depth diffusion model denoises it in one step.

Key results

• On RGB-D-D at native 192x144 input, AdaDS reaches RMSE 0.0607 and MAE 0.0332, beating the second-best PriorDA by 7.33% RMSE.
• On ScanNet with 16x downsampling plus degradations, AdaDS improves over the second-best PromptDA by 16.91% on average RMSE.
• Ablation at 8x shows the learned noise sampling module and adaptive timestep rule are each important; random timestep raises RMSE from 0.0597 to 0.0768 on RGB-D-D.

Surprising findings

• Direct regression of the refined latent is fragile on TOFDSR, performing worse than the calibration stage alone.

Limitations

• Robustness to very large upsampling or severe degradation depends on the calibration stage's uncertainty estimate, and performance drops sharply for large tau.
• Training relies on synthetic Hypersim data with simulated degradations, so real sensor degradations are only approximated.

Experimental setup

Datasets: Hypersim RGB-D-D TOFDSR NYUv2 ScanNet

Benchmarks: RGB-D-D TOFDSR NYUv2 ScanNet

Metrics: RMSE MAE delta 1.05 accuracy

Baselines: Depth Anything V2 Marigold-LCM SGNet C2PD DuCos PromptDA OMNI-DC Marigold-DC PriorDA

Builds on: Marigold Latent Consistency Model Depth Anything V2 DDPM

Closest papers by meaning

• SharpDepth: Sharpening Metric Depth Predictions Using Diffusion Distillation (CVPR 2025 · similarity 0.89)
• Zero-shot Depth Completion via Test-time Alignment with Affine-invariant Depth Prior (AAAI 2025 · similarity 0.88)
• DepthFM: Fast Generative Monocular Depth Estimation with Flow Matching (AAAI 2025 · similarity 0.87)
• Prompting Depth Anything for 4K Resolution Accurate Metric Depth Estimation (CVPR 2025 · similarity 0.86)
• Depth Prompting for Sensor-Agnostic Depth Estimation (CVPR 2024 · similarity 0.86)
• Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (ICLR 2025 · similarity 0.86)