Training Custom object detection network using tensor-flow object detection API?

Question

I was just wondering if some one could provide a nice tutorial on how to use the Recent tensor-flow object detection API to train custom network say like VGG-16? (Just USE the VGG-16, VGG-19, Inception-v3 etc as a fixed feature extractor in the Faster RCNN implementation).

From their Documentation I know it can be done. As, am an amateur in AI, I am unsure how to move forward.

P.S: This could really help some keras fans (like me :P) who have A trained CNN in keras. They could just make a .pb file out of it and insert it as a fixed feature detector/extractor in the Tensorflow Object detection API. Voila!!! Now Keras is integrated with the Tensorflow's object detection API

Answer 1

I have found a hack to integrate keras with tensorflow object detection API. This hack works if you have trained the keras classification model with tensorflow backend. If above is the case you can extend the classification model to a object detection model by first converting the keras checkpoint to a tensorflow checkpoint then in the object detection API write new feature extractor layers using tf.keras (Keras is now part of core tensorflow starting from version 1.4). you can simply copy paste your layer definitions and pass the path to the converted checkpoint in the config file and it will work.