Custom human training data
ghavidelsahar opened this issue · 0 comments
ghavidelsahar commented
Hi there, I have a question about data preparation for custom data. This dataset is labeled in a way that canonical pose and human pose has large rotation and translation differences. Do you think this affects model training? because so far I was not able to get any decent output on this dataset
Do you have any suggestion besed on the following results? I have used zju config file and all default parameters without modification, however, loss_c does not converge properly.
Obtained output for val_ind: