Text-Image-Augmentation About Joint Training

Thanks for your work, but I am wondering how to joint training, which is mentioned in the paper (Algorithm 1 Joint Learning Scheme).

Mar 23 '20 02:03 DongfeiJi

I have two related questions as well:

Is the "moving state" S the deltas between src_pts and dst_pts (src_pts - dst_pts) which contains 2(N+1) points?
Is the loss function for joint learning cross-entropy loss? If so, what is the representation of GT for this loss function? Any explanation/idea/suggestion/discussion is welcomed :)

Sep 21 '20 19:09 yanfjz

@DongfeiJi This version is the default random augmentation. We guess it is sufficient for practical use.

Dec 02 '20 07:12 Canjie-Luo

@yanfjz

The "moving state" denotes the moving directions.
Yes, we use cross-entropy loss. The GT is the moving state that increases difficulty.

Dec 02 '20 07:12 Canjie-Luo

@Canjie-Luo I read the paper Learn to Augment: Joint Data Augmentation and Network Optimization for Text Recognition.
and I'm so interested in joint training (Algorithm 1 Joint Learning Scheme)，can I ask for the whole code of paper?There is something I can't understand and I want to retrain it .I'll appreciate it very much .Here's my email [email protected]

Jan 30 '21 13:01 zdz1997

@Canjie-Luo

I have some questions about the paper.

Output size of the agent network is 2x(N+1)x2x2. Isn't just two coordinates enough to predict the direction of each point?
In Algorithm 1, randomly select one point in S and switch to the opposite direction to make S prime. Does this mean choosing one moving state in a mini-batch?

Mar 17 '21 08:03 DaeHwanGi

I read the paper Learn to Augment: Joint Data Augmentation and Network Optimization for Text Recognition. and I'm so interested in joint training (Algorithm 1 Joint Learning Scheme)，can I ask for the whole code of paper? my email: [email protected]