Generalization capability about msn-point-cloud-completion

[Colin97]

Sorry, I have not tried that before. It's an interesting point, but I guess it's less likely for the model to transfer to a completely novel categories.

[MartinSmeyer]

Thanks a lot for your answer!

What do you think about training jointly on multiple categories? In real-world/robotic scenarios, your method would be useful, but the number of categories is usually very large whereas shapes are often shared.

Nice work by the way!

[MartinSmeyer]

Just as a reference, this paper has a nice discussion about it:
https://arxiv.org/pdf/2006.15427.pdf

[Colin97]

Our model is trained on multiple (8) categories, but I guess generalizability may be hard for our model. Yes, there are some recent works taking this issue into account and my recent work on learning-based mesh reconstruction also possesses strong generalizability.