[x] Ensure that images without annotations are excluded from the
training/eval data sets.
[x] Resize images before creating the TF record files.
[x] Try to use a DNN with higher mAP, we've failed to train with
faster_rcnn_inception_resnet_v2_atrous_coco, but got the following error:
https://stackoverflow.com/q/46135528/814224
[x] This script crops an image to just the content of the text, test it:
https://github.com/danvk/oldnyc/blob/master/ocr/tess/crop_morphology.py
[ ] Create a standalone package with the training data for bib detection.
[ ] Replace bib detection model by faster_rcnn_nas (?).