The question about side-refinement.

[JacobianTang]

Hi:
I have a question about side-refinement.Traing time,you regress the anchor's relative offset to bounding.Testing time,what do you do ? We can get the text proposal according the defalut anchor,
then ,add the offset (compute with current anchor) to text proposal ,yes or no? Maybe ,the side-refinement only be used on the training time,to get better text proposal.On the test time,nothing should be done.

[tianzhi0549]

@JacobianTang in test phase, the text proposals are firstly connected into text line, and then the side-refinement is applied to the text proposals on both ends.

[LearnerInGithub]

@tianzhi0549 same question as @JacobianTang, how could I recovery the full four coordinates of a fine-grained bbox from a vertical regression target(center_y, height)?