Getting Bounding Boxes with highest confidence during prediction

Question

I am using TensorFlow Object Recognition API. Normally, what developers do is to setup a training pipeline, provide some checkpoints or tfrecords to start the training while monitoring the performances on TensorBoard. This is what I did and right now I can see all the predicted bounding boxes on Tensorboard, that varies according tot he number of iteration. But what if I need to get these bounding boxes? Is there any line of code that given an image it returns the predicted bounding boxes?

Answer 1

If you are using the sess.run(...) command for inference, it'll return an python dictionary object (for example called output_dict). It contains everything, that the model is supposed to return, e.g output_dict['detection_boxes'][0], output_dict['detection_scores'][0] and output_dict['detection_classes'][0]. You can iterate through this dictionary in the common 'pythonic' way. For example:

box_index = 0
for box in output_dict['detection_boxes'][0]:
   current_box = box
   current_class_id = output_dict['detection_classes'][0][box_index]
   current_score = output_dict['detection_scores'][0][box_idx]

   # Do something with box

   box_index += 1

Edit: As mentioned above, you can use the jupyter notebook to calculate a 'out of the box' inference with a frozen graph. For production use, have a look at Tensoflow Serve.