search

bharat-b7 / LoopReg

91 stars 10 forks source link

Pretrained weight (Not naked) #3

Open tyluann opened 3 years ago

tyluann commented 3 years ago

I've noticed that you have provided the model for naked input. Would you kindly provide the pre-trained model for non-naked input?

Thanks in advance.

MoyGcc commented 3 years ago

I'm also interested in that. Will this be released in the future?

bharat-b7 commented 3 years ago

Hi, I'm a bit tied till ICCV deadline. But in my experiments the pre-trained naked model works almost as good as the dressed one for dressed data. This works due to two reasons:

  1. Correspondences between the undressed and dressed shape are similar (except for loose clothing).
  2. LoopReg optimizes the correspondences and model parameters on the fly, so even when correspondences are not very accurate, LoopReg auto adjusts.

Let me know if this doesn't work.

Luciano07 commented 3 years ago

Hi, I was testing Loopreg with pre-trained naked model, using this scan as input: scan_test.zip But the input / output looks like this: Screenshot from 2021-03-29 16-31-26

I think the clothes make the SMPL model wider than it should, maybe with pre-trained for non-naked model the result could be improved.

Thanks for your responses!

chris-hndz commented 3 years ago

Hello @bharat-b7 !

Thanks for sharing your great proyect. I'm also using LoopReg and I had similar results as @Luciano07

It should be great if you could provide the weights for a full-dressed (non-naked) input. It will be very helpfull for our research.

Thanks in advance!

wangfudong commented 3 years ago

@Luciano07 Hi, I downloaded your obj file in "scan_test.zip" and found that it looks like image

Luciano07 commented 3 years ago

@wangfudong my mistake, here is the correct one: scan_correction.zip

wangfudong commented 3 years ago

@Luciano07 Thanks. Moreover, have you successfully run the released code? I find that both the provided readme and code can not be directly ran for neither training nor testing/inferring/validation.

Luciano07 commented 3 years ago

I ran it, i created a virtual environment with the requirements and modified some paths and it works. whats the error do you have?

wangfudong commented 3 years ago

@Luciano07 Aha，I also rewrote some functions/calsses like make_data_split/dataloader and it works now. The predicted smpl model (without offsets) of your data 'scan_correction.zip' looks similer with your results, next I will try to rewrite these released codes such that it can also fit scans with SMPL+D models.

Luciano07 commented 3 years ago

@wangfudong, What do you mean by 'offsets'?, Are you removing some offset in SMPL?. Using SMPL+D sounds good, please let me know if you have any results with these.