search

shunsukesaito / SCANimate

This repository contains the code for the paper "SCANimate: Weakly Supervised Learning of Skinned Clothed Avatar Networks"
https://scanimate.is.tue.mpg.de/
Other
273 stars 31 forks source link

The demo training produces unsatisfactory results. How to obtain better results? #13

Closed HalleyJiang closed 2 years ago

HalleyJiang commented 2 years ago

After training with the data of example_03375_shortlong, I obtain unsatisfactory results. There are many noise points. Further, the left upper leg is weird. Is it because the number of frames (16 only) is not enough? Do you have suggestions to obtain better results?

Screenshot from 2021-12-22 19-24-09

JinlongYANG commented 2 years ago

There is a chance the model is strongly overfitted. If you follow strictly as the demo in the repository, train the model and test it on the AIST sequence, this is likely to happen. You just need to set a smaller number of max epochs for igr_net training in the config file: num_epoch_sdf: 20000 -> 1000 in ./configs/example.yaml This is the result we have, trained with 1000 epochs and tested on AIST: https://youtu.be/59z-oemRdas The 20000 epochs are just to make sure you can fully reproduce the training sequence.