How to use SynthText for training?

[lufo816]

Text in SynthText is oriented and its label has 4 points, how to use these data for training?

[MhLiao]

@lufo816 use the maximum rectangle(xmin,ymin,xmax,ymax) as the bounding box

[lufo816]

@MhLiao Understood, thanks a lot!

[vrbala]

@MhLiao I am trying to use SynthText to fine tune the model. However, I am not sure how to preprocess the data such that it is suitable to use with TextBoxes. Specifically, I couldn't figure out the schema of data in lmdb files. Could you kindly point me at some relevant references? Thanks in advance.

[MhLiao]

@vrbala The schema of data in lmdb files is similar to the SSD. You can use the scripts of SSD to generate the lmdb data.

[vrbala]

@MhLiao Thanks so much.

[HelloTobe]

@MhLiao Hi, the SynthText datas are in decimal form, should i use the round function to make them into integers?
What pictures in SynthText did you use for training, validation or testing respectively? (I mean i don't know what pictures should be categorized into training, also is validation and testing.)
Thanks in advance.

[MhLiao]

@HelloTobe yes. I used 800k images for training, 58750 images for validation.

[HelloTobe]

Thanks for your reply!
By the way, i'm wondering how long you finish your training.

[HelloTobe]

@MhLiao Hi, thanks for your reply!

Another problem is that some bounding boxes exceed image boundary(some coordinates are too large or negative), could you give me some adviece to solve this problem?

By the way, i'm wondering how long you finish your training.

[lomograb]

@HelloTobe You can enable CUDA or cuDNN (GPU MODE) for quick result of training time

[offlaner]

@MhLiao @lufo816 For training on synthText are we supposed to generate txt files in the form of :
xmin ymin xmax ymax "text"
instead of :
Xcentre Ycentre w h "text"

[github14343]

@HelloTobe Have you solved the problem " bounding box exceeds image boundary." Thanks for help

[HelloTobe]

@github14343 I probably just clip the coordinate x such that 0 <= x <= H (or W).

[github14343]

@HelloTobe Thanks!