AI Paper Atlas

Research map / 3D vision / Depth and stereo

PatchRefiner: Leveraging Synthetic Data for Real-Domain High-Resolution Monocular Metric Depth Estimation

Zhenyu Li, Shariq Farooq Bhat, Peter Wonka

ECCV 2024 · 3D vision › Depth and stereo · approach: Foundation models for monocular geometry · 13 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
reformulate high-resolution depth as residual refinement of a frozen coarse model (PatchRefiner vs PatchFusion)ZoeDepth · UnrealStereo4KRMS: 1.088 -> 0.892Improved
predict residual on coarse depth instead of direct depth (ablation)PatchRefiner · UnrealStereo4KRMS: 0.925 -> 0.892Improved
add more feature levels (1 to 6) to refiner decoder (ablation)PatchRefiner · UnrealStereo4KRMS: 0.926 -> 0.892Improved
add DSD loss with pseudo labels from synthetic teacherPatchRefiner · CityScapesBoundary Recall: 40.92 -> 48.78Improved
mix synthetic and real data during trainingPatchRefiner · CityScapesdelta1 (%): 95.418 -> 89.108Hurt
fine-tune on real data after synthetic pretrainingPatchRefiner · CityScapesBoundary F1: 23.04 -> 23.68No clear effect
supervise pseudo labels with silog loss instead of rank+ssi (ablation)PatchRefiner · CityScapesdelta1 (%): 95.359 -> 82.550Hurt
restrict silog pseudo-label loss to regions lacking GT (masking) (ablation)PatchRefiner · CityScapesdelta1 (%): 82.550 -> 81.425No clear effect
combine rank and ssi losses instead of either alone (ablation)PatchRefiner · CityScapesBoundary Recall: 47.81 -> 48.78Improved
increase DSD loss weight lambda from 0.1 toward 1 (ablation)PatchRefiner · CityScapesBoundary Recall: 48.78 -> 58.48Mixed
use DSD loss versus standard fine-tuningPatchRefiner · ScanNet++RMS: 0.268 -> 0.268No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy