Out-of-distribution monocular depth estimation with local invariant regression

Monocular Depth Estimation (MDE) aims to produce a high-quality depth map from a single RGB image. Recent advancements in deep learning have led to notable improvements in MDE through end-to-end manner. However, most existing approaches primarily focus on MDE using clean input data, neglecting the challenges posed by Out-of-Distribution (OoD) corrupted data in the real-world scenarios. This oversight limits the generalization and robustness of advanced MDE models, resulting in unpredictable errors. To tackle this issue, in this paper, we propose an OoD MDE model that employs local invariant regression. Specifically, we introduce a novel HOG (Histogram of Oriented Gradients) consistency loss function, which utilizes local invariant feature regression to capture robust local geometric structures and enhance robustness to the common corruptions in a self-supervised manner. Additionally, we present a novel Cascaded Iterative Enhancement network (CIE-Depth), designed to accurately and robustly predict monocular depth maps through adaptive hierarchical interaction in a coarse-to-fine manner. Extensive experiments demonstrate that our method achieves highly competitive performance against state-of-the-art techniques on the KITTI, NYUv2, SUN RGBD and underwater FLSea datasets. Further evaluations highlight the robustness of our approach when applied to the corrupted KITTI-C and NYUv2-C datasets.