Reconstruction loss for variational inference (PointNet case)

[space1panda]

Hello,

Could you please elaborate on the absence of reconstruction loss in the objective function? The BCE loss is between gt occupancy and predicted occupancy, it is not a reconstruction of geometry - so how exactly we guarantee that encoded latents represent pointclouds? Thanks for the answer