Research map / 3D vision / Depth and stereo

Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth

NeurIPS 2026 · 3D vision › Depth and stereo · approach: Self-supervised monocular depth · 0 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 Depth2Pose, which scores monocular depth estimators by the accuracy of relative pose estimated with depth-aware RePoseD solvers, needing only camera poses rather than ground-truth depth, and introduces the D2P dataset of 24 challenging scenes.

Problem

Monocular depth estimators are evaluated only by pixel-wise depth accuracy against dense ground truth, which is costly to obtain and does not reflect how useful depth is for downstream geometric tasks or how well it generalizes to out-of-distribution scenes.

Approach

Replace ground-truth depth with ground-truth relative camera poses (from COLMAP) and evaluate MDEs by the pose error of depth-aware relative pose solvers (RePoseD via PoseLib) that consume predicted depth at matched keypoints, yielding a task-driven, geometry-weighted depth quality proxy measured with mAA(10°).

Key results

• On standard benchmarks, MoGeV2-L achieves the top mAA(10°) under the H estimator (82.65) and rank 1 in delta1 (90.59), while the no-depth 5-point baseline reaches 84.01 on the same setup.
• On D2P-Vegetation, UniK3D-L ranks first (79.95 mAA under H) while MoGeV2-L ranks 10th of 12, showing the ranking on standard benchmarks does not transfer.
• Using MDEs does not yet improve pose accuracy over the no-depth baseline on the evaluated datasets; the GT-depth upper bound reaches 90.70 (H) on standard benchmarks, indicating large room for improvement.

Surprising findings

• Depth-aware pose estimation with current MDEs does not beat the no-depth 5-point baseline on the evaluated datasets.
• Rankings of MDEs on standard benchmarks do not transfer to the D2P dataset; the best standard-benchmark method (MoGeV2) falls to tenth on D2P-Vegetation.

Limitations

• Only applicable to scenes where accurate camera poses can be obtained (e.g., via SfM).
• Relative pose has scale ambiguity, so metric depth accuracy cannot be assessed.
• Only regions with correspondences are evaluated, so accurate depth in textureless or mirror regions cannot be distinguished from absence of accuracy there.

Experimental setup

Datasets: ETH3D LaMAR ScanNet++ Sintel D2P

Benchmarks: ETH3D LaMAR ScanNet++ Sintel D2P

Metrics: mAA(10°) AbsRel delta1

Baselines: MoGeV2 MoGeV1 Metric3DV2 UniK3D UniDepthV2 Pi3 DepthAnythingV3 MapAnything DepthPro DepthAnythingV2 5-point solver (No Depth + B)

Builds on: RePoseD PoseLib LO-RANSAC COLMAP LoMa

Closest papers by meaning

• Relative Pose Estimation through Affine Corrections of Monocular Depth Priors (CVPR 2025 · similarity 0.91)
• DepthMaster: Unified Monocular Depth Estimation for Perspective and Panoramic Images (NeurIPS 2026 · similarity 0.90)
• Seurat: From Moving Points to Depth (CVPR 2025 · similarity 0.90)
• A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · similarity 0.90)
• Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (ICLR 2025 · similarity 0.90)
• DME: Unveiling the Bias for Better Generalized Monocular Depth Estimation (AAAI 2024 · similarity 0.90)