Object Localisation(From Scratch)

Create and build a model(from scratch) to detect a simple bounding box i.e build a regression head on Oxford IIT Data set

Objective of this notebook

• The purpose of this notebook is to build a model to detect a simple bounding box
• Details of the problem statement , data set , summary of the code/solution , sample output/Prediction from the program and final result of the project are listed in the sections to follow.

Problem Statement

Build a model to detect a single Bounding Box on a pet image

Data Description:

• Data Set :Oxford IIT Pet Data
• Link : https://www.robots.ox.ac.uk/~vgg/data/pets/

Summary of the Solution/Code:

The code aims at building a simple bounding box to detect the animal's face

• We begin by building a utility class to to read the input images ,convert them to a structured format for further processing and model building and pickle the created input data files Code @ Read_& Structure_Input.ipynb
• We then do an EDA on the data and check various statistics on the images such as total no of images , whether classes are balanced and display a few images with their bounding boxes for visulaization . Code @ EDA.ipynb
• We will then we will pre-process the input data to make it compatible for model building wherein we first do a Train-Test split and create 2 subsets (70:30 split), convert all images to tensors for both subsets,resize all the images to one pre-defined size, convert bounding boxes accordingly ,visualize data and see that everything looks good. Code @ Pre-Processing_Data.ipynb
• Finally we build the model, train the model , log the results and run an inference.We will build a model that is capable of detecting a bounding box on the image -i.e object localisation.We will try 3 different variants of the model. One variant is we will build a Convolution Neural Network from SCRATCH ,train it & capture results.The second variant is we will build an image classifier converted into an object detector i.e a pre-trained MobileNet base with a basic top layer re-built, re-trained and log results.The third is we will then try a model with pre-trained MobileNet base and a slightly more sophisticated top layer re-built, re-trained and log results .We compare and contrast score , pick the best model and run an inference. Code @ Train_And_Inference.ipynb

Sample Ouput/Prediction :

Here is a sample Ouput image from the program/model

Result

• 3 Different models were tried to build the regression head/ bounding box
• A model with pre-trained mobile Net base with a basic top layer reconstructed gave the best results of validation IOU of .96 in 30 epochs
• A model with pre-trained mobile Net base with a more complex top layer reconstructed gave us a result of .8372 in 30 epochs . Further training could possible give us hihger numbers
• A 6 layer deep CNN model built from scratch gave us a result of .5295 in 30 epochs . Further training could possible give us higher numbers.
• The best model is model no 1( model with pre-trained mobile Net base with a basic top layer reconstructed)

References & Guidance

• PGP GL/UAT study material
• https://colab.research.google.com/github/zaidalyafeai/Notebooks/blob/master/Localizer.ipynb