FaceAlignment_Multi-task
trying to reproduce the result of TCDCN, but failed.
I think the model is slightly overfitting, train--1e-4 and val--4e-4. regularzation does not make improvement.
the results are poor when come to images cropped by myself(out of test set)
some hyperparameters are hard to choose~~~ may be this is another reason of the poor result~~~
data augmentation
random translation, zoom, rotation, noise, blur, flip, constrast jitter used
small jitter is good. check out face images after augmentation
there are some tricks get verified in this experiment...