Basic queries for Object Detection using deep learning

Question

I have recently started working on the Object detection using deep learning.

I have understood the basic steps of object detection are:

Input Image -> Object proposals -> Feature Extraction -> Training.

I have successfully implemented my own code up to the Object Proposal extraction. Now I want to use deep learning here from that point. For the input to deep learning I have fixed number of object proposals (i.e. 800 proposals from one image) and I need a labeled data. For example: I have a one class dataset. 1) Aircraft, Images number 1-90 are aircraft. and the above number of images have 800 aircraft.
Q: Here, I want to ask that how can I make the labels for the dataset with this configuration of class? According to the tutorials and the Matlab implemented examples I have seen, they used the images and the training labels as the input to the CNN. As I have training images I want to know how can i make training labels.

The input to the CNN network will be proposals (just a cropped regions) around 500 proposals. I want the output is to be if any of the proposals is matched as aircraft then it will appear a bounding box on the big image that this is a aircraft.

Input image.

Answer 1

The input to your CNN is a small cropped image, and the output will be a prediction at the label (either "airplane" or "not-airplane"). The training set needs to contain example input-output pairs. So, you need your training set to contain many small images (of the size of a single crop), with a label indicating whether it should be classified as "airplane" or "not-airplane".

How do you do that? You take many large images where you know where the bounding box of the airplanes in each is. For each large image, you crop it around an airplane and add that to the training set, labelled as "airplane". You also crop it somewhere else (in a place not intersecting any of the airplane bounding boxes) and that to the training set, labelled as "not-airplane". You can do the letter step many times, taking many different crops. Repeat this for many large images. Use the resulting training set to train your neural network.