Need help for training

I used my own datasets and continued training on the released checkpoint.

I noticed that there were obvious jagged rectangular boundaries in the generated depth images. Could you please tell me what the reason for this might be?

I use MSE loss for training. Does it matter to use MAE loss instead?