Training for DDad dataset

[liucsg]

There's a lot of noise, when training on ddad datasets. I want to know the impact when depth is sparse , which comes from single-frame lidar. More importantly, how to deal with it?