DeepLearning with Pytorch - Projects | Portfolio Antoine Cougny

This repository will host the projects I did for the OpenCV.org courses on Deeplearning with Pytorch. Each of them were achieved in Fall 2020.

Link to the repository here .

Project 1 - Classification from Scratch - 100%#

This project consisted in implementing the different steps of the training pipeline used to train an image classifier from scratch. The aim of the project was then to classify pictures of dogs, cats and pandas in their respective class. Simple task for a human, much harder for a machine.

It consists of two Jupyter notebooks, the first one focusing on:

Step 1 - Understand your problem
Step 2A - Get the data
Step 2B - Explore & Understand your data
Step 2C - Create a sample data from the dataset
Step 3 - Data Preparation
Step 4 - Train a simple model on sample data and check the pipeline before proceeding to train the full network

The second part focuses on defining the actual model, training it and improving it to reach the minimal validation accuracy of 85%.

The training graph looks like this:

training graph

The confusion matrix:

confusion_matrix

An inference on a sample of data from the training set:

sample_inference

Project 3 - Object Detection - 100%#

Wikipedia states that

Electronic toll collection (ETC) is a wireless system to automatically collect the usage fee or toll charged to vehicles using toll roads, HOV lanes, toll bridges, and toll tunnels.

ANPR (Automatic Number Plate Recognition) is one of the challenges encountered by the ETC. This project consisted in solving this problem.

Using the Vehicle registration plate public dataset and the detectron2 framework developed by Facebook AI Research, I trained a neural network to detect license plates.

Here are a sample of the result:

first_car second_car third_car

The final task consisted in processing the following video and inferring each frames in order to detect and locate the license plate on the different frames.