mks0601 / Hand4Whole_RELEASE

Official PyTorch implementation of "Accurate 3D Hand Pose Estimation for Whole-Body 3D Human Mesh Estimation", CVPRW 2022 (Oral.)
MIT License
315 stars 31 forks source link

run demo.py get bad results #74

Open HonglinChu opened 1 year ago

HonglinChu commented 1 year ago

Thank you for your work,when i run demo.py( main branch), i get bad result as follow: image

Then i checkout branch Pose2Pose, and run face_demo.py, i get right result as follow: image

Is there something wrong in main branch ?

HonglinChu commented 1 year ago

run demo.py in main branch, get bad mesh in face image

mks0601 commented 1 year ago

Did you use the same flame files?

HonglinChu commented 1 year ago

I have downloaded them from the official website, and is there other version about flame files? Would you mind sharing me smplx files and flames files by google cloud?

image
mks0601 commented 1 year ago

If Pose2Pose works well on your side, could you use the same flame file as Pose2Pose of yours for the Hand4Whole demo?

HonglinChu commented 1 year ago

Hand4Whole demo (main branch) and FaceNet(Pose2Pose branch) are using different files, it seems impossible to use the same file image

mks0601 commented 1 year ago

Oh sorry. could you try with this? https://drive.google.com/drive/folders/1t7fnNnDUK0DZnr73SSYpz8zyiNne_8Uh?usp=share_link

HonglinChu commented 1 year ago

Hi,if i set fixed expr and jaw_pose value, i can get good result as follow: image image

I also save the face bbox image, it is normal: image

image

So, the input of self.face_regressor is right, but the output expr is error. Maybe, there is a problem with self.face_regressor ?

mks0601 commented 1 year ago

I've just ran the demo by cloning this repo without any modification and got good result like below. Please check what you modified

image

HonglinChu commented 1 year ago

Thank you,the problem is solved when i reinstall the cudatoolkit=11.x