yfeng95 / scarf Goto Github PK

Thanks for your great work!
I wonder if we could extract the clothes mesh seperately, and turn them to T-pose?
Besides, the reconstructed clothes are one-piece, is there any way to seperate the clothes into upper body and lower body?

Hello, the link you provided in fetch_data.sh for the scarf_utilities.zip is not accessible. Could you please share it again?

Thank you.

Hi, your model can do fully body textured ?

Can you please upload the pre-trained weights as well as a script to run inference on our own videos @yfeng95.

Hello, is the data set withdrawn?

Hello. First of all, thank you for revealing your awesome works with open source.

I am curious how can we obtain 'full_pose' parameter, which consists of 5533 size tensor per frame, in your given demo pose file 'pixie_radioactive.pkl'.

I think this can be driven by PIXIE, but I don't know how can we get it.

PIXIE parameters from the preprocessing have different shapes compared to the given one.

Thank you.

Hi, thanks for the great work! I have 2 questions:

1. Why is camera intrinsics and pose not used in sampling rays?
2. Where is the customized volumetric rendering?

Hello, the website of more data can not be opened. Could you please upload the data again?

The pre-trained model of cloth-segmentation used in the code is no longer valid, so I would like to ask if you have a backup? thanks!

Thanks for the excellent work!

When I try to train the stage 0 from scratch (without using the pretrained model provided in the code), I got all black images.
The code works well when I try to train it with the pretrained model.
What might be the reason for this?

zz

Thanks for your great work.

When I train the hybrid stage, the losses become Nan, as following grpah illustrates.

And also, the process of optimization is very slow(200 steps about 30 minutes.)

When starting to train the hybrid model, the loss function appears to be anomalous and the images show all black images.
It seems to be caused by the ID-MRF loss function.