Research map / 3D vision / Depth and stereo

DepthMaster: Unified Monocular Depth Estimation for Perspective and Panoramic Images

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 DepthMaster, a unified metric depth estimator that canonicalizes panoramas into overlapping 95° perspective cubemap patches, injects virtual camera pose/intrinsics as tokens, and uses a correspondence consistency loss to stitch patches while keeping a standard DINOv2 ViT-L backbone.

Problem

Metric monocular depth estimation that generalizes across both narrow-FoV perspective images and 360° panoramas, where panoramic models are specialized and suffer from scarce metric panoramic training data and ERP distortion.

Approach

Decompose panoramas into overlapping perspective patches with virtual camera conditioning and enforce cross-patch consistency via a feature-level correspondence loss (CCL) instead of ad-hoc cube-convolution or fusion modules, so perspective priors transfer directly to panoramas.

Key results

• On panoramic benchmarks, DepthMaster reduces average metric-depth AbsRel from 11.79 (DAP) to 7.25 and scale-invariant AbsRel from 6.05 (DA2) to 4.77.
• On 10 perspective datasets, DepthMaster attains the best average rank (1.31) vs. 2.22 for MoGe 2, with affine-invariant point map avg Rel 6.18 and delta1 93.9%.
• Removing CCL raises panoramic AbsRel from 6.31 to 7.41 while perspective performance is nearly unchanged.

Surprising findings

• Trained with only one panoramic dataset (Structured3D), it beats UniK3D on Matterport3D even though UniK3D was trained on that dataset.
• Panoramic-only training degrades perspective performance severely (AbsRel 7.15 to 16.34), showing perspective data is needed for cross-domain prior transfer.

Limitations

• Struggles with highly abstract out-of-distribution imagery such as ink wash paintings.
• Inference is slower than DAP (531 ms vs 427 ms) due to six patches per panorama.
• Metric depth on some perspective datasets (e.g., KITTI, Spring) lags specialists in the metric depth setting.

Experimental setup

Datasets: Structured3D Hypersim A2D2 MVS-Synth GTA-SfM UrbanSyn TartanAir IRS ARKitScenes MegaDepth BlendedMVS KenBurns

Benchmarks: Stanford2D3D Matterport3D PanoSUNCG NYUv2 KITTI ETH3D iBims-1 GSO Sintel DDAD DIODE Spring

Metrics: AbsRel RMSE delta1 delta2 average rank

Baselines: UniK3D DAC DA360 DAP DA2 MoGe 2 UniDepth V2 Depth Pro Metric3D V2 Depth Anything V2

Builds on: DINOv2 Depth Anything 3 MoGe DPT

Closest papers by meaning

• Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.90)
• Unified Panoramic Geometry Estimation via Multi-View Foundation Models (NeurIPS 2026 · similarity 0.90)
• Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (ICLR 2025 · similarity 0.89)
• RSA: Resolving Scale Ambiguities in Monocular Depth Estimators through Language Descriptions (NeurIPS 2024 · similarity 0.89)
• SharpDepth: Sharpening Metric Depth Predictions Using Diffusion Distillation (CVPR 2025 · similarity 0.89)
• UniDepth: Universal Monocular Metric Depth Estimation (CVPR 2024 · similarity 0.89)