Review for NeurIPS paper: UCSG-NET- Unsupervised Discovering of Constructive Solid Geometry Tree

Additional Feedback: I tend to accept this paper because it demonstrates a very promising probability - that it is now possible to learn a relatively complex shape representation (CSG) that are often used in actual production settings. Granted, the novelty is slightly limited (the training protocol is similar to earlier works in implicit shape generation/reconstruction and the representation (CSG) is also widely used), and the results are not super convincing (will explain below); however, I still feel the idea is interesting enough and will easily sparkle future works in similar directions. I am not completely positive because I am uncertain if the method leads to a dead end - as there are infinitely many possible CSG trees for a given shape, an unsupervised method might never be able to learn something that is truly usable, even after many improvements over the current method. Additional comments: -Could the authors explain the motivation of adopting a bottom-up process that groups primitive to form final outputs? It seems to me that it would be much harder for the network to understand shape decomposition in this way, since it is trying to select form a large set of primitives and find which ones can be grouped to form something that look similar to the input, rather than directly thinking about how to decompose the input shape (in a top-down way).