PAINFULLY slow to do prediction with kurapan/EAST implementation

[jhaggle]

I was trying out the code by running the evaluation script with the following command:

python3 eval.py --gpu_list=0 --test_data_path=example_images/ --model_path=tmp/icdar2015_east_resnet50/EAST_IC15+13_model.h5 --output_dir=tmp/icdar2015_east_resnet50/eval/

The images used are the ones provided by Kurapan here:
https://github.com/kurapan/EAST/tree/master/examples

On my CPU it took about 10-20 seconds for the test images. On my gpu it took about 0.25 seconds per images.
If I instead used my own images it took about one second per image. In both cases it is really slow.

Why is it so slow? Is there any way I can speed it up?