Will you release the trained model？

[ReeseL]

Hi:

Thanks for your contribution. I am intreasted in human reconstruction. Will you release the trained model SVR for SingleView Human Reconstruction？

Thanks!

[jchibane]

Hi,

you're welcome! I uploaded you a trained model here https://drive.google.com/file/d/1ccaYI-O4jmAOCw19bPKumTZcfYbnjBlt/view?usp=sharing

You will have to adapt the generation script such that it 1) loads your single view human data as input 2) loads the checkpoint for the SVR model

Best Julian

[ReeseL]

Hi，

Thanks for your reply! I have a question: checkpoint_epoch_6.tar could not be unzipped. Error: tar: This does not look like a tar archive tar: Skipping to next header tar: Exiting with failure status due to previous errors

If it is convenient for you, could you please upload it again?

Thanks!

[jchibane]

Hi,

you don't need to unzip this one, you can directly load it into pytorch in a similar manner to https://github.com/jchibane/if-net/blob/master/models/generation.py#L101

I might add support of pre-trained models to the repository in a later stage, in a few weeks.

Best, Julian

[ReeseL]

Hi,

Thanks for your answer! I get it.

Best, ReeseL

[ReeseL]

Hi,

I loaded my single view human data as input and generated a full model. But, the result is not right. I am not sure the parameters I entered are correct.

I hope you can give me some advice. Thanks a lot!

parameter: python generate.py -pointcloud -pc_samples 3000 -std_dev 0.2 0.015 -res 256 -m SVR -checkpoint 6 -batch_points 100000

input:

result:

[jchibane]

Maybe an issue with the scaling? Our human data on the hight (Y) axis goes from -0.7 to 1 approximately, so rescaling might help. Also, I would suggest looking into the voxelized point cloud input to the net. Visualization will show if that's what you expect it to be.