Good and bad news...
At this point I have reached a dead end - greater dataset and more learning steps dont give better hands. Pose estimation works amazingly well but hands are like total random. I'm going to try some more but probably I will change aproach - maybe model only based on hands, maybe additional preprocessor creating hands depth maps etc.
RELEASE:
For now you will need to use external preprocessor avalaible HERE!
Models are available HERE!