Boosting Depth Estimation for Self-Driving in a Self-Supervised Framework via Improved Pose Network

Depth estimation is a critical component of self-driving vehicles, enabling accurate scene understanding, obstacle detection, and precise localization. Improving the performance of depth estimation networks without increasing computational cost is highly advantageous for autonomous driving systems. In this article, we propose to enhance depth estimation by improving the pose network in a self-supervised framework. Unlike conventional pose networks, our approach preserves more detailed spatial information by integrating multi-scale features and normalized coordinates. This improved spatial awareness allows for more accurate depth predictions. Comprehensive evaluations on the KITTI and Make3D datasets show that our method yields a 2-7% improvement in the absolute relative error (abs_rel) metric. Furthermore, on the KITTI odometry dataset, our approach demonstrates competitive performance, with relative translational error ($t_{rel}$) of $6.11$ and $7.21$, and relative rotational error ($r_{rel}$) of $1.12$ and $2.05$ for sequences 9 and 10, respectively.