This is an implementation of YunYang1994's excellent TensorFlow implementation of YOLOv3, found here, with specific utilities and modifications for dealing with detecting the location of license plates in images.
This was tested on a dataset of approximately 400 images of Hong Kong license plates, with transfer learning with a pretrained model trained on the COCO dataset, and 95% of the bounding boxes were correct, even with a confidence cut-off of 20%.
It works pretty well in low light AND with skewed license plates too:
