Closed liuweie closed 1 year ago
Have you tried the demo data and cfg? It seems the problem of your custom keypoints, are they of shape (3,137)?
In fact,the image showed above is exactly demo data of oliver,and oliver's keypoints shape is (122, 2):
import numpy as np
path = 'D:\\Pose2Img\\data\\Oliver\\keypoints\\Auto_Lending_-_Last_Week_Tonight_with_John_Oliver_HBO-4U2eDJnwz_s_58_000_0_00_05.405405.npy'
npy = np.load(path)
print(npy.shape)
and the result:
(122, 2)
please pull the code
Oh!I tried keypoints of shape(3,137), problem solved now, thanks for your patience !
May I ask what are the specific meanings of the two parameters kp_var_root
and mask_sigma_perp
? Is there any explanation in the paper?And here are some of my img result using vis_hyperparams.py ,does this prove that this is a good parameter setting?
The two parameters are used to control the size of gaussian map. The keypoint gaussian maps are used to indicate the target pose. The per-part mask gaussian maps are used to fetch the source features and they will be refined in the sub module. You can refer to Synthesizing Images of Humans in Unseen Poses for more details. It seems a proper hyperparams setting. If you produce poor results, you can augment your custom training data with more diverse poses to bridge the gap between train and test distribution. For further discussing, you could contact me by email zhiyh@shanghaitech.edu.cn.
Thanks very much!
hello,when I use
vis_hyperparams.py
to computekp_var_root
andmask_sigma_perp
,the following error ocurred:I've been debugging for a long time and still can't solve it. Can you help me to see what the problem is? .