Figure 1. Previous methods (a) directly fuse the RGB information
aligned with the LR depth, while our method (b) focuses more on
modeling the degradation representation of the LR depth to pro-
vide targeted guidance for HR depth recovery.