search

wangyida / softpool

SoftPoolNet: Shape Descriptor for Point Cloud Completion and Classification - ECCV 2020 oral
Apache License 2.0
74 stars 5 forks source link

About the code #6

Closed brandon9838 closed 3 years ago

brandon9838 commented 3 years ago

Hi, I'm trying to reproduce the result by training softpoolnet, but it seems that train.py, model.py are a mixture of softpoolnet, grnet, and pointnet. The latter 2 works are not included in the architecture reported in your paper.

If I only want to reproduce the completion result of softpoolnet, can I comment out output3, output4, part_regions, emd3, emd4 in train.py(line 115), and the associated networks in model.py?

By the way, are there tensorflow codes for softpoolnet in the tensorflow folder?

Thanks for replying!

wangyida commented 3 years ago

Hi yep you can comment out those output produced for grnet and foldingnet, together with architectures in model.py. We have softpool operaters in tensorflow folder, the softpoolnet architecture is not identical to our paper in TensorFlow's implementation.

Notice that the emd loss function there should be either earth moving distance or chamfer distance