Pose Networks Unveiled: Bridging the Gap for Monocular Depth Perception

Abstract

Depth estimation is essential in Augmented Reality applications,enabling realistic object placement, scene understanding, spatialmapping, interaction, and environment awareness. This paper proposesa method to enhance depth model performance without increasinginference costs by improving the pose network in a selfsupervisedlearning setup. In particular, we enrich spatial informationin the pose network by incorporating features from differentscales and normalized coordinates. It is experimentally shown onthe KITTI dataset that our approach achieves a 2-7% improvementin the abs rel metric when compared to baseline techniques.