Learning Joint 2D-3D Representations for Depth Completion | IEEE Conference Publication | IEEE Xplore