How can we predict dense points from sparse points?

We first triangulate 2D sparse points to 3D and then do a sparse-to-dense prediction using a multi-layer perceptron. We then reproject 3D points back to 2D to get 2D dense annotations.