Traning on new dataset and pre-training on smaller SynthText Dataset

[shervin-gohari]

Hi,

We are group of students working on generating 3D engine synthetic datasets. Our aim is to train and evaluate our dataset using TextFuseNet. We currently have three issues which we hope we can get some help with.

First, our generator produces images and annotations on the ICDAR 2015 format meaning we get bounding boxes with 4 coordinate pairs starting from upper left and going clockwise. Do we need to rewrite these to the same format as COCO-text? Are there any scripts available for doing this?

Secondly, we note that in the paper its described that the training on ICDAR2015 is done with a weakly supervised approach. Is this the method used when executing the described method under "Train a new model"?

Finally, we want to pre-train on a subset of SynthText. Can we do this by running the method under "Train a new model"? Do we need the segmentation maps for our subset of SynthText? In that case, how are the segmentation maps obtained?