Pretrained weight (Not naked)

[tyluann]

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]

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

[bharat-b7]

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]

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:

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]

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]

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

[Luciano07]

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

[wangfudong]

@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]

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]

@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]

@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.