Geometric Distillation from Rectified Stereo: Leveraging Epipolar Cues for Monocular Depth

2026-07-17 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

Monocular depth foundation models face challenges with metric depth estimation in diverse environments due to inherent scale ambiguity, leading to misaligned scale predictions despite accurate relative geometry. While multi-view foundation models leverage cross-view cues for robust scene geometry and consistent scale, these advantages typically diminish during single-image inference without explicit geometric constraints. To address this, a novel framework is proposed that transfers scale-aware geometric priors from multi-view models into monocular depth foundation models. This framework introduces Epipolar Distillation (EpiDistill), which utilizes Rectified Stereo Tokens. EpiDistill enables single-view prediction models to maintain epipolar attention patterns and geometric consistency without requiring multi-view inputs during inference. Experimental results show significant improvements in zero-shot metric depth estimation, particularly on challenging datasets such as ETH3D and DIODE, where precise scale alignment is crucial. The approach is model-agnostic, consistently enhancing the performance of state-of-the-art ViT-based models, including UniDepthV2 and DepthPro.

Key takeaway

For Computer Vision Engineers developing monocular depth systems, this framework offers a solution to the persistent challenge of scale ambiguity. You should consider integrating Epipolar Distillation to transfer robust geometric priors from multi-view models, significantly improving zero-shot metric depth estimation. This approach enhances existing ViT-based models like UniDepthV2 and DepthPro, ensuring more accurate scale alignment in diverse environments without requiring multi-view inputs at inference.

Key insights

Epipolar Distillation transfers multi-view geometric priors to monocular depth models, improving scale-aware metric depth without multi-view inference.

Principles

• Scale ambiguity hinders monocular metric depth.
• Multi-view models offer robust scale-aware geometry.
• Epipolar cues enable geometric consistency transfer.

Method

Epipolar Distillation (EpiDistill) uses Rectified Stereo Tokens to transfer scale-aware geometric priors from multi-view models to monocular depth foundation models, enabling single-view prediction to retain epipolar attention patterns and geometric consistency.

In practice

• Apply EpiDistill to ViT-based depth models.
• Improve zero-shot metric depth on ETH3D/DIODE.
• Enhance UniDepthV2 and DepthPro performance.

Topics

• Monocular Depth Estimation
• Multi-view Geometry
• Epipolar Distillation
• Foundation Models
• Computer Vision
• Scale Ambiguity

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, Machine Learning Engineer

Related on AIssential

• Visual Place Recognition in Forests with Depth-Aware Distillation — Depth-aware distillation enhances forest Visual Place Recognition by integrating geometric cues into DINOv2, improving robustness.
• Mesh-Aware Epipolar Matching for Multi-View Multi-Person 3D Pose Estimation in Basketball — Dense mesh geometry significantly enhances training-free multi-view 3D pose estimation in complex sports environments.
• Honey, I Shrunk the Arc de Triomphe! — A new dataset and method mitigate scale-collapse in monocular depth estimation for distant, unconstrained scenes.
• The 3D Mirage: Probing and Taming 3D Hallucinations — Monocular depth models hallucinate 3D structures from 2D illusions, requiring targeted benchmarks and mitigation.
• Multi-scale interaction network for stereo image super-resolution — A multi-scale interaction network enhances stereo image super-resolution by optimizing intra-view and cross-view information exploitation.
• Learning Geometric Representations from Videos for Spatial Intelligent Multimodal Large Language Models — GeoVR enables MLLMs to learn 3D spatial intelligence from 2D videos by distilling geometric knowledge via multi-objective learning.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.