Breaking the Horizontal Prior: From Long-Tailed Orientation Bias to Roll-Robust Monocular Depth Estimation

2026-08-01

AI Summary

This work addresses a critical yet previously overlooked limitation in monocular depth estimation: the significant performance degradation under slight camera roll due to a pervasive “horizontal prior” bias in training data. The study is the first to identify and formally name this issue, proposing an inference-architecture-preserving training strategy termed Invariant Depth Constraint (ID-Constraint). This approach enhances robustness to roll perturbations by jointly regularizing the depth backbone through auxiliary geometric and spatial reasoning tasks within a multi-task learning framework. Extensive experiments across five benchmark datasets and four distinct roll configurations demonstrate that the proposed method substantially improves both accuracy and stability of depth estimation under non-level camera orientations, without modifying the underlying inference architecture.

Abstract

Despite recent advances in Monocular Depth Estimation, state-of-the-art depth foundation models remain vulnerable to robustness issues. Particularly, even slight camera rolls can result in substantial degradation in depth estimations. We attribute this problem to a previously overlooked phenomenon, termed the Horizontal Prior, which is a manifestation of long-tailed distribution bias: most training images are captured in approximately horizontal orientations due to human visual preferences and photographic habits. While intuitive remedies such as re-balanced data augmentation and horizon leveling provide partial improvements, they fail to fully address the issue. In this paper, we introduce Invariant Depth Constraint (ID-Constraint), a training-time supervision strategy that improves roll robustness by fine-tuning and jointly regularizing the depth backbone with a series of geometric and spatial reasoning tasks. These auxiliary objectives encourage the backbone to learn rotation-stable, depth-relevant representations, while the auxiliary prediction heads are discarded after training, leaving the original inference architecture unchanged. Extensive experiments on five benchmark datasets across four roll settings demonstrate the effectiveness of the proposed method.

Problem

• Monocular Depth Estimation
• Roll Robustness
• Horizontal Prior
• Orientation Bias
• Long-Tailed Distribution

Innovation

• Invariant Depth Constraint
• roll robustness
• monocular depth estimation
• horizontal prior
• rotation-stable representation