Toward Explainable Monocular Depth Estimation

Loading...
Thumbnail Image

Authors

Hosseini, S.M. Hossein MohammadHossein

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

Modern monocular depth estimation systems achieve impressive accuracy through large-scale training, yet operate as black boxes, posing challenges for safety-critical certification and generalization to novel scenarios. This thesis quantifies the role of three principles of human depth perception: ground theory, which anchors depth on recovery of the ground surface; perceptual organization, which partitions the image into semantic regions to locally regularize estimates; and linear perspective, which exploits geometric regularities from the projection of parallel lines. By integrating these cues, we construct an explainable monocular depth estimation framework that, despite its simplicity, rivals opaque deep networks in specific contexts. We further show that fusion with our system reduces absolute relative error by up to 33.4% in indoor ground regions and by up to 10.7% overall on SYNS-Patches. These findings suggest that large-scale training does not fully encode fundamental geometric priors, and that explicit geometric modeling remains essential for robust depth inference.

Description

Keywords

Computer science, Electrical engineering

Citation