NNNet: New Normal Guided Depth Completion From Sparse LiDAR Data and Single Color Image

In this paper, we propose new normal guided depth completion from sparse LiDAR data and single color image, named NNNet. Sparse often uses maps as a constraint for model training. However, direct construction of map the image causes lot noise in reduces performance. Thus, use an intermediate to promote fusion multi-modal features. We generate it network The is generated by converting input into grayscale constructing map, replacing Z channel with original depth, finally adding mask. Based on construct end-to-end NNNet its corresponding image. consists two branches. one branch generates while other constructs dense predicted map. branches fully merge features through skip connection. loss function, L2 ensure that plays restrictive role. Finally, refining spatial propagation network. Experimental results show provides effective constraints completion. Moreover, achieves 724.14 terms RMSE outperforms most current state-of-the-art methods.