Anchor Points for Monocular Depth Estimation Under Severe Rolling Shutter Ego-motion
Jacob Carter, Vaishnav Ramesh, Md Jahidul Islam, and Sanjeev J. Koppal
Rolling shutter (RS) cameras are widely deployed in robotic systems due to their low cost and power efficiency. However, robot ego-motion induces significant RS distortions that degrade geometric consistency and severely challenge visual perception. In this work, we address monocular depth estimation under large rolling shutter distortions caused by ego-motion. We introduce the concept of an anchor point, which establishes a consistent temporal and geometric reference between rolling shutter and global shutter observations. This formulation enables principled adaptation and fine-tuning of deep foundation models for depth estimation under RS distortions. To systematically evaluate the problem, we present a new rolling shutter dataset featuring large motion-induced geometric distortions representative of mobile robotic tasks. Our experimental analysis shows that anchor point selection is a primary factor governing depth accuracy under RS effects. We further characterize anchor configurations for RS-GS data capture and identify stable regimes that yield accurate depth estimates. Using a representative two-stage depth estimation pipeline, we conduct extensive ablation studies demonstrating that anchor-based modeling significantly improves performance, outperforming recent RS correction baselines under strong ego-motion. Finally, we validate our approach on a legged robot platform, demonstrating robust monocular depth estimation in the presence of substantial rolling shutter distortions in under 400 ms.