search

paschalidoud / superquadric_parsing

Code for "Superquadrics Revisited: Learning 3D Shape Parsing beyond Cuboids", CVPR 2019
https://superquadrics.com
Other
135 stars 33 forks source link

Configuration for reproduction of ShapeNet v2 Experiment #7

Closed ktertikas closed 2 years ago

ktertikas commented 3 years ago

Hello @paschalidoud , thank you very much for the well-written codebase and paper! I wanted to ask some things related to the reproduction of the ShapeNet Core v2 dataset experiments.

In particular, in the example of the README file, a suggested way of training the model on all chair data examples of ShapeNet is as follows:

./train_network.py ~/data/03001627/ /tmp/ --use_sq --lr 1e-4 --n_primitives 20 --train_with_bernoulli --dataset_type shapenet_v2 --use_chamfer

From what I can understand reading through the code, the loss used when running the above command does not include the parsimony loss mentioned in the paper. Could you let me know how to include this loss term?

I can see there is indeed a parsimony loss in the codebase (link is here), but this is not the same equation as the one mentioned in the paper.

Thank you for the very well-structured repo once again!

mahmoudEltaher commented 2 years ago

Hi @ktertikas

I have same question, did you get any update

thanks