GIFT: Geometry-Invariant Fine-Tuning for Non-Lambertian Monocular Depth Estimation

2026-08-03

AI Summary

This work addresses the challenge of depth hallucination in monocular depth foundation models when encountering non-Lambertian surfaces—such as specular or transparent objects—where reflections or transmissions mislead depth estimation. To mitigate this issue without requiring ground-truth depth labels, the authors propose a parameter-efficient, unsupervised fine-tuning framework. By capturing multi-view RGB images under fixed geometry to observe appearance variations, the method leverages geometric invariance constraints to suppress depth artifacts in non-Lambertian regions. This approach represents the first integration of geometric invariance into unsupervised fine-tuning for monocular depth estimation. Evaluated on both a newly curated benchmark and real-world datasets, it significantly improves depth accuracy for specular and transparent objects while preserving strong generalization performance on conventional Lambertian surfaces.

Abstract

Monocular depth foundation models, benefiting from large-scale synthetic training data, have demonstrated strong generalization. However, they often hallucinate depth on non-Lambertian surfaces, estimating reflected content in mirrors or transmitted content behind glass rather than the physical surface itself. Adapting these models with real-world data is challenging because conventional depth sensors are also unreliable in such regions. We observe that while the appearance of a non-Lambertian surface varies with its reflected or transmitted environment, its underlying geometry remains unchanged. Based on this observation, we propose GIFT (Geometry-Invariant Fine-Tuning), a parameter-efficient post-training framework that requires no measured depth labels. We collect groups of RGB images under controlled appearance changes while keeping the camera and target geometry fixed. GIFT exploits geometric invariance across these observations to suppress non-Lambertian depth hallucinations while retaining general depth estimation capability. We further construct a controlled benchmark that evaluates non-Lambertian depth recovery, robustness to appearance changes, and performance retention in other regions. Experiments on our benchmark and an independent real-world dataset demonstrate that GIFT improves depth prediction for mirrors and transparent objects while largely preserving the base model's performance, providing a practical and low-cost approach for adapting monocular depth foundation models to non-Lambertian scenes.

Problem

• monocular depth estimation
• non-Lambertian surfaces
• depth hallucination
• mirror
• transparent objects

Innovation

• Geometry-Invariant Fine-Tuning
• Non-Lambertian Surfaces
• Monocular Depth Estimation
• Depth Hallucination
• Parameter-Efficient Adaptation