dunbar12138 / pix2pix3D

pix2pix3D: Generating 3D Objects from 2D User Inputs
http://www.cs.cmu.edu/~pix2pix3D/
MIT License
1.65k stars 131 forks source link

what is the detail of limitations of second "dataset's distribution" #24

Open justinday123 opened 1 year ago

justinday123 commented 1 year ago

what is the actual limitation?

dunbar12138 commented 1 year ago

Hi,

The limitation is that our model cannot generate out-of-distribution results, even if such input is given. Fig 16 in our arXiv paper is an example.

justinday123 commented 1 year ago

thanks. If there is no 'green cat' in dataset, than your model can't make '3d green cat?'. Or, your model trains about one class in one time?(ex: if your model trains about AFHQ, your model cannot generate about all the class in AFHQ in once?) Also, the reason why your model cannot generate is because of EG3D?

Thank you for your comment

dunbar12138 commented 1 year ago

If there is no 'green cat' in the training set, our model won't be able to generate it. Yes, I think it is a limitation for most of the GAN-based methods.