search

LizhenWangT / NormalGAN

NormalGAN: Learning Detailed 3D Human from a Single RGB-D Image (ECCV 2020)
72 stars 13 forks source link

How to get the depth picture #5

Open sunjc0306 opened 3 years ago

sunjc0306 commented 3 years ago

Hi! Thank you very much for your excellent work. I've got great reconstruction. However, I have a question how to get the depth map corresponding to the RGB picture?

LizhenWangT commented 3 years ago

You should first calibrate your depth camera and get the internal and external parameters. Then, please refer to Registration in (https://github.com/r9y9/pylibfreenect2) as an example.

sunjc0306 commented 3 years ago

Thanks. I read your paper carefully. However, I didn't find three discriminators (Fdb, Fcb, Fdf) in the code. Are there detailed descriptions of these. In addition, I would wonder to ask if there is a plan for open source training code.

Luciano07 commented 3 years ago

Hi, I'm using 'pylibfreenect2' to capture rgb and depth images from Kinect v2, containing background and body information. I tested NormalGAN with these images, but it fails in erosion function. I looked in 'datasets/testdata/' and the images only contain body pixels. Seems like that i have to apply a body mask to use NormalGAN correctly. Which method do you use for the segment body in color and depth images?

Thanks for amazing work!

LizhenWangT commented 3 years ago

Hi, I'm using 'pylibfreenect2' to capture rgb and depth images from Kinect v2, containing background and body information. I tested NormalGAN with these images, but it fails in erosion function. I looked in 'datasets/testdata/' and the images only contain body pixels. Seems like that i have to apply a body mask to use NormalGAN correctly. Which method do you use for the segment body in color and depth images?

Thanks for amazing work!

Thank you! Actually, we directly cut the body part using two thresholds in depth map in live demos. This may lead to bad results around the feet part (depth map also performs bad in this area). If you do not care about this area, that's enough.

LizhenWangT commented 3 years ago

Thanks. I read your paper carefully. However, I didn't find three discriminators (Fdb, Fcb, Fdf) in the code. Are there detailed descriptions of these. In addition, I would wonder to ask if there is a plan for open source training code.

As we can not distribute our dataset due to commercial reasons, we think it's hard to generate a similar dataset (which needs hundreds of 3D human models). So we don't have the plan to update the training code.