search

skhu101 / GauHuman

Code for our CVPR'2024 paper "GauHuman: Articulated Gaussian Splatting from Monocular Human Videos"
Other
339 stars 29 forks source link

How to render people in canonical/T pose #49

Closed JiatengLiu closed 4 weeks ago

JiatengLiu commented 1 month ago

Hi, Shoukang. I want to know how to render the people in the T-pose. I have tried to skip the process coarse_deform_c2source, but the rendering results is all zeros, how can I fix it?

JiatengLiu commented 1 month ago

I commented on the code below, but the result was not a face facing the camera like below.

can_pts = torch.matmul(A_fif[..., :3, :3], query_pts[..., None]).squeeze(-1)
smpl_src_pts = can_pts + A_fif[..., :3, 3]

image

skhu101 commented 1 month ago

Hi, thanks for your interest. Do you visualize rendered images in T-pose?

JiatengLiu commented 1 month ago

yes, besides I want to render people towards cameras with T-pose.

skhu101 commented 1 month ago

Hi, you can set the target pose to be T-pose to have a debug. You can get rendered results under T-pose by skipping the coarse_deform_c2source function.

JiatengLiu commented 1 month ago

Hi, Shoukang, I have followed your guide to skip the coarse_deform_c2source function, but the results is not what I wished. Results is like below. image What I want to get is like below. image Or can we turn the problem into how to remove the steering of the human body during rendering so that the human body is always facing the camera

skhu101 commented 1 month ago

Hi, if you hope to use the same camera as the target image, you can change the target pose to be T-pose while keeping other parameters (shape, camera parameters) the same. The other solution is to set up a lookat camera by referring to https://github.com/skhu101/GauHuman/issues/29.

JiatengLiu commented 1 month ago

Thanks for your reply! I will try later.

Best!

At 2024-10-21 22:34:08, "Shoukang Hu" @.***> wrote:

Hi, if you hope to use the same camera as the target image, you can change the target pose to be T-pose while keeping other parameters (shape, camera parameters) the same. The other solution is to set up a lookat camera by referring to #29.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>