First of all, your tutorial is a great explanation about object detection! Well done!
Lately, I have been trying to replicate your results but using TensorFlow instead of Keras. However, my model is not converging as yours. I realized your Keras version is an old one (e.g., border_mode of Convolutional2D is not available in current documentation), so I took care to use the same parameters you used.
You can see the complete code here: https://gist.github.com/arnaldog12/2f6e8910f56a73e95be682143193682e
Could you help me trying to understand why is it happening, please?