Object Localisation for Car Identification

Objective of this notebook

• The purpose of this notebook is to build a Deep learning based car identification model
• Lets note here that we term this as object localisation as the data set at hand has one object per image & therefore we can aim at building an object localisation model instead of a conventional object detection model
• 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.
• Side Note : This notebook is part of our capstone project submission files done during our PGP AI/ML

Problem Statement

Computer Vision can be used to automate supervision and generate action appropriate action trigger if the event is predicted from the image of interest. In this particular use case, we will be exploring how a car moving on the road can be easily identified by a camera by identifying the make of the car, type, color, number plates, etc.This could be very useful for monitoring criminal activity, identifying fake license plates, etc. It could also be useful in several other use-cases like identifying popularity of certain car types/make in certain geographies etc

Data Description:

• Data Set:Stanford Car Data Set hosted by Kaggle
• The Cars dataset contains 16,185 images of 196 classes of cars. The data is split into 8,144 training images and 8,041 testing images, where each class has been split roughly in a 50-50 split. Classes are typically at the level of Make, Model, Year, e.g. 2012 Tesla Model S or 2012 BMW M3 coup
• Train Images: Consists of real images of cars as per the make and year of the car.
• Test Images: Consists of real images of cars as per the make and year of the car.
• Train Annotation: Consists of bounding box region for training images.
• Test Annotation: Consists of bounding box region for testing images

Domain:

Automotive Surveillance

Summary of the Solution/Code:

The code aims at building a Custom CNN model.

• We begin by doing an Exploratory Data analyses and Visualisation on the images , refer python worksheet /EDA/EDA_And_Preprocessing.ipynb

• We then do the required pre-processing for the data to make it compatible with the model to be built.We also perform the data splits because we will not be using 16,000 images due to hardware constraints , we will perform a Train Data Split Size=6500 Records and Validation Data Split Size=1629 Records

• We then move on to Model building which will be a simple keras classifier model that has been converted to an object detector. Essentially it is a convolution neural network(in this case with a base skeleton of MobileNet)which is pre-trained on a ImageNet DataSet who’s top layer have been replaced by two heads , namely a regression head and a classification head. This can be thought of as two parallel networks , one which does classification and the other that calculates the bounding box , each having its own loss function respectively. The combined loss of these functions is what the network strives to reduce with each epoch during the training.

  Architecture Diagram of the above model:

• We then do "Hyper parameter tuning" which involves going through multiple iterations while building the model to try different options for parameters such as base skeleton n/w, top layers, optimizer and image size

• The best model resulting from the above iterative process is chosen as the model refer python worksheet /Custom Baseline Model/Final_Custom_Baseline_Model.ipynb

Sample Ouput/Prediction :

Here is a sample Ouput image to the program/model alongside the ground truth

Result

• This model was able to obtain an accuracy of 0.70, IOU of 0.82 and mAp score of 0.592 .Refer python worksheet /Custom Baseline Model/Calculate_mAp.ipynb for mAp score calculation

References & Guidance

• PGP GL/UAT study material
• Capstone Mentor guidance

Extension

• For this data set , we had the luxury of building a custom object localisation model because a single object exists in each image of this data set & therefore we should start with the simplest model possible - a custom object localisation model
• Now as an extension to this worksheet , we could try conventional object detection models like YOLO and SSD to see whether they were able to give us better scores and log the results. -It would be interesting to see if using complex models like YOLO or SSD on this data set just ended up being an overkill