Research map / 3D vision / Depth and stereo

SharpDepth: Sharpening Metric Depth Predictions Using Diffusion Distillation

CVPR 2025 · 3D vision › Depth and stereo · approach: Self-supervised monocular depth · 4 citations (OpenAlex)

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 SharpDepth, a diffusion-based sharpener that refines metric depth from UniDepth using noise-aware gating, SDS distillation from Lotus, and a noise-aware reconstruction loss, trained on ~90k unlabeled real images without ground-truth depth.

Problem

Monocular metric depth estimators are accurate but over-smooth, while diffusion-based depth models give sharp boundaries but only relative depth and are trained on synthetic data with a domain gap.

Approach

Refine a metric depth prediction with a diffusion model initialized from Lotus, using a difference map between metric and affine-invariant depth to selectively add noise (noise-aware gating) and to weight both an SDS loss (detail) and a reconstruction loss (metric accuracy), enabling ground-truth-free training on real images.

Key results

• SharpDepth reaches KITTI delta1 97.3 and A.Rel 0.06, close to UniDepth (97.9 / 0.05), and avoids the large accuracy drop of relative-depth models on metric benchmarks.
• On Sintel, SharpDepth's PDBE completion error is 36.2 versus 113.3 for UniDepth, and on UnrealStereo4K it is 61.5 versus 257.3.
• Ablation: replacing the difference-map input with Gaussian noise raises KITTI RMSE from 2.37 to 3.98; removing the reconstruction loss drops KITTI delta1 to 0.843.

Measured effects reported in the paper

Surprising findings

• Without the SDS loss, the model mirrors UniDepth and achieves the best KITTI RMSE (2.28) but much worse Sintel completion error (112.5), indicating a trade-off between accuracy and detail.

Limitations

• Sharpness gains come with some accuracy loss on certain benchmarks (for example ETH3D and Diode are weaker than UniDepth-level results in the reported table).
• Depends on the quality of the input metric depth model and the teacher diffusion model for its difference map and distillation targets.

Experimental setup

Datasets: Pandaset Waymo ArgoVerse2 ARKit Taskonomy ScanNetv2 KITTI NYUv2 ETH3D Diode Booster nuScenes

Benchmarks: KITTI NYUv2 ETH3D Diode Booster nuScenes iBims Sintel UnrealStereo4K Spring

Metrics: delta1 A.Rel RMSE Pseudo Depth Boundary Error (accuracy and completion) Rank

Baselines: UniDepth Metric3Dv2 ZoeDepth DepthAnythingV2-Metric PatchRefiner Marigold Lotus BetterDepth UniDepth-aligned Lotus

Builds on: Lotus Marigold UniDepth DreamFusion SwiftBrush

Closest papers by meaning

• Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (ICLR 2025 · similarity 0.90)
• UniDepth: Universal Monocular Metric Depth Estimation (CVPR 2024 · similarity 0.90)
• DepthFM: Fast Generative Monocular Depth Estimation with Flow Matching (AAAI 2025 · similarity 0.90)
• DepthMaster: Unified Monocular Depth Estimation for Perspective and Panoramic Images (NeurIPS 2026 · similarity 0.89)
• Metric Depth Estimation from Arbitrarily Degraded Low-Resolution Depth Prompts (NeurIPS 2026 · similarity 0.89)
• BetterDepth: Plug-and-Play Diffusion Refiner for Zero-Shot Monocular Depth Estimation (NeurIPS 2024 · similarity 0.88)