search

boycehbz / MvSMPLfitting

A multi-view SMPL fitting based on smplify-x
MIT License
251 stars 28 forks source link

how to fitting other dataset #1

Open ccsvd opened 3 years ago

ccsvd commented 3 years ago

hi, it looks very good for the demo result and thank you release the code. now i have detected the h36m data by using hrnet to get 2d point.(coco format like you use,have 4 cameras); then runing fitting,but the final loss is large(~119198),and result looks bad. so the code is suited to other dataset? if i want label other data, how should i adjust the code? could you please give me some advices? Thanks.

boycehbz commented 3 years ago

This code is suitable for other multi-view datasets. However, you need to provide the 2D keypoints and calibrated camera parameters for input. You can try to set the cam_param in cfg_files/fit_smpl.yaml to the camera parameters of H36M.

ccsvd commented 3 years ago

yes,i get the input data: data.zip and get the bad result. can you try?

boycehbz commented 3 years ago

It seems your 2D keypoints are not correct. You can check the 2D keypoints and camera parameters again. 1

ccsvd commented 3 years ago

i have check the 2d keypoint and camera again, it looks fine: green color point is gt image and the fitting result is not correct! i try another dataset(mpi_inf_3dhp),and the result looks prety good: image

maybe the h36m camera parameters have some problem,i will check this later.

ccsvd commented 3 years ago

hi, can we use the male or female smpl model to replace the neutral? if i want use it,how should i do? thanks.

boycehbz commented 3 years ago

set the gender flag in cfg_files/fit_smpl.yaml to male/female

Kartik-Teotia commented 3 years ago

i have check the 2d keypoint and camera again, it looks fine: green color point is gt image and the fitting result is not correct! i try another dataset(mpi_inf_3dhp),and the result looks prety good: image

maybe the h36m camera parameters have some problem,i will check this later.

Hey, how exactly did you obtain the camera parameters for mpi_inf_3dhp?

boycehbz commented 3 years ago

i have check the 2d keypoint and camera again, it looks fine: green color point is gt image and the fitting result is not correct! i try another dataset(mpi_inf_3dhp),and the result looks prety good: image maybe the h36m camera parameters have some problem,i will check this later.

Hey, how exactly did you obtain the camera parameters for mpi_inf_3dhp?

3DHP provides the ground-truth camera parameters. You can find it in the downloaded files.

Kartik-Teotia commented 3 years ago

i have check the 2d keypoint and camera again, it looks fine: green color point is gt image and the fitting result is not correct! i try another dataset(mpi_inf_3dhp),and the result looks prety good: image maybe the h36m camera parameters have some problem,i will check this later.

Hey, how exactly did you obtain the camera parameters for mpi_inf_3dhp?

3DHP provides the ground-truth camera parameters. You can find it in the downloaded files.

Hey, thanks for the reply. I also wanted to ask what exact command are you using with openpose to generate the keypoints? The one I am using (COCO) is giving 18 keypoint locations instead of 17 .

boycehbz commented 3 years ago

i have check the 2d keypoint and camera again, it looks fine: green color point is gt image and the fitting result is not correct! i try another dataset(mpi_inf_3dhp),and the result looks prety good: image maybe the h36m camera parameters have some problem,i will check this later.

Hey, how exactly did you obtain the camera parameters for mpi_inf_3dhp?

3DHP provides the ground-truth camera parameters. You can find it in the downloaded files.

Hey, thanks for the reply. I also wanted to ask what exact command are you using with openpose to generate the keypoints? The one I am using (COCO) is giving 18 keypoint locations instead of 17 .

The joint order in this demo is COCO17 image You can map the keypoints from openpose to this order and save the results like data/keypoints/0000/Camera00/00001_keypoints.json