Suri Zeeshan Khan

Abstract

Self-supervised monocular depth estimation approaches suffer not only fromscale ambiguity but also infer temporally inconsistent depth maps w.r.t. scale.While disambiguating scale during training is not possible without some kind ofground truth supervision, having scale consistent depth predictions would makeit possible to calculate scale once during inference as a post-processing stepand use it over-time. With this as a goal, a set of temporal consistency lossesthat minimize pose inconsistencies over time are introduced. Evaluations showthat introducing these constraints not only reduces depth inconsistencies butalso improves the baseline performance of depth and ego-motion prediction.