Pose Networks Unveiled: Bridging the Gap for Monocular Depth Perception

Depth estimation is essential in Augmented Reality applications, enabling realistic object placement, scene understanding, spatial mapping, interaction, and environment awareness. This paper proposes a method to enhance depth model performance without increasing inference costs by improving the pose network in a self-supervised learning setup. In particular, we enrich spatial information in the pose network by incorporating features from different scales and normalized coordinates. It is experimentally shown on the KITTI dataset that our approach achieves a 2-7% improvement in the abs_rel metric when compared to baseline techniques.