| KEYWORD |
Generative Adversarial Networks for Domain Adaptation between synthetic and real images
Thesis in external company
keywords GAN, GANS, GENERATIVE ADVERSARIAL NETWORKS
Reference persons BARTOLOMEO MONTRUCCHIO
External reference persons Enrico Busto enrico.busto@add-for.com
Research Groups GR-09 - GRAphics and INtelligent Systems - GRAINS
Thesis type RESEARCH
Description Learning from as little human supervision as possible is a major challenge in Machine Learning. In the context of computer vision, Deep Learning is a class of supervised learning algorithms that require a great amount of human-labeled images in order to be trained [1]. An opportunity to reduce the needed amount of human labelling is offered by synthetic dataset, where the labelling procedure comes at almost no cost. The problem with the introduction of synthetic datasets is the domain gap with real images ones.
Generative Adversarial Networks (GANs) [2, 3] are a class of deep neural networks able to generate synthetic data with the same distribution of a target dataset. They have been applied to very different areas with excellent results, from the generation of realistic face images to 3d points cloud [4].
The aim of this thesis is to study the possibility of using GAN to cover the domain gap between real and synthetic images [5]. The objective is to develop an algorithm that minimises the amount of labeled real data needed to train a target deep neural network while obtaining the best classification accuracy.
SEE ALSO:
[1] Deep Learning
http://neuralnetworksanddeeplearning.com/
[2] Generative Adversarial Networks
https://arxiv.org/abs/1406.2661
[3] Wasserstein GAN
https://arxiv.org/pdf/1701.07875.pdf
[4] The GAN Zoo
https://github.com/hindupuravinash/the-gan-zoo
[5] SPIGAN: Privileged Adversarial Learning from Simulation
https://openreview.net/pdf?id=rkxoNnC5FQ
Required skills Proficiency in at least one programming language (Python, Lua, Matlab, C++, Java);
Basic knowledge of machine learning, in particular, supervised learning;
Good knowledge of linear algebra.
Deadline 01/06/2020
PROPONI LA TUA CANDIDATURA