Here, I have trained two CNN architectures :
-
One predicts bounding boxes of a single class - Airplanes as shown below
Notebook :bounding_box_regressor.ipynb
-
The other is fine tuned for predicting class labels as well as boxes for 5 different classes chosen at random from the CALTECH101 dataset.
The link to the dataset, necessary preprocessing can be found out inmulticlass_bounding_box.ipynb
The two output branches, one for classification and other for bounding box regression, improve the overall efficiency of the task cause being the use of a better loss function.
Tutorial : PyImageSearch