Time permitting, you will enhance the results by designing a network that analyses short video fragments.

The software will be developed in Python using Jupyter Notebook development kit. For deep learning you will use the PyTorch framework.

This is an advanced course, the knowledge of Python is assumed.
(erwartete Hörerzahl original: 12, fixe Veranstaltung: nein)

Title:

Lab Course Machine Learning in Signal Processing

Credits: 2,5

Prerequisites

Knowledge of Python programming language is required. Basic theoretical knowledge in machine learning is assumed: consider taking the Machine Learning in Signal Processing (MLSIP) course in the same semester.

Contents

Imagine a car driving on an autobahn in an automatic mode. Among other things, the car needs to steer itself to keep driving in it's own lane. To accomplish this, the central problem is to detect the road-lane markings. These are the white solid or dashed lines that are drawn on each side of the lane. The standard modern approach to solve this type of problems is to take a large dataset of labled examples and train a deep neural network model to accomplish the task. This is how car and pedestrian detection algorithms are developed. The difficulty with the road-lane markings is that there is no labled dataset of them and creating such dataset would cost millions of dollars.
In this lab course we will solve this problem using transfer learning and mathematical modeling:

• Create cartoon-like artificial images of a road with known locations for the lane markings.

• Train deep neural network on these artificial images with heavy data augmentations that mimic real-world images.

• Create a dataset of unlabeled real-life videos by downloading and organizing examples from youtube.

• Create a machine learning pipeline for working with these videos efficiently.

• Apply the neural network that has been trained on artificial data to the real world videos.

• Analyze the quality of results produced by the network.

• Use mathematical modeling to correct the outputs of the network.

• Retrain the network on the dataset composed of the corrected outputs.

• Measure and analyze the quality of the results.
  

The software will be written in Python using JupyterLab development framework. Access to modern GPU servers will be provided. This is an intensive research-level course; the result of the course might be the creation of state-of-the-art lane detection system for self-driving cars.