references

• paper
• a github repo where i get train images and its processed labels

difference of mine from the paper

• no side refinement
• no recurrent network

data source

baidu cloud drive

dependency

• tensorflow(1.12.0)
• PIL
• numpy

scripts

• train.py #for training
• eval.py #for evaluation
• config.py #all configs are congregated here
• model.py #build the model
• tf_vgg_pretrained.py #convert a keras vgg16 model to tensorflow

use

edit eval.py at line 19 to put your image path,then run python ./eval.py.You will see result image pops up with bounding boxes drawn on it.

demo