Object Detection

This repository contains a simple project for the implementation of an object detection system, using bounding boxes. The project was implemented using Python programming language and specifically the Keras framework.

Dataset

To run this code, dataset Caltech 101 should be downloaded. The next step is the extraction of the contains of airplanes folder inside the path dataset/images and the contains of Airplanes_Side_2 inside the path dataset/annotations. The structure of the project has to be as following:

dataset/
dataset/images
dataset/annotations
output/
test/
bounding_box.py

Execution

To compile and run the source code, you must have Python downloaded and execute:

python bounding_box.py

Code Explanation

The code makes use of Tensorflow and Keras libraries in order to train a Regression Model to predict 4 values. These 4 values correspond to the corners of a bounding box, which indicates the object inside our input image. The implementation of the model was based on Transfer Learning, by loading VGG-16 without the head FC layer and then freezing all of its layers, so they will not be updated during training process. This network was then connected to a series of Dense layers that end to an output layer with 4 neurons, each one corresponds to a point of the predicted bounding box.

The optimizer used was Adam, while the loss function was MSE, which is regularly used to Regression problems. After 10 epochs, the performance of the model was already converged, as we can see in the graph below:

Testing

To test the performance of our model we can see the predicted bounding boxes to some random images, sampled for test. An example can be seen below: