Excavating the Potential Capacity of Self-Supervised Monocular Depth Estimation