Description

Description
Deep Learning excels when huge volumes of labelled data are available. However, when the data lack labels or only a small portion of them are labelled, techniques like Self-supervised and Semi-supervised learning are introduced. Self-supervised learning enables the network by learning features on a completely unlabelled dataset with proxy-tasks, while Semi-supervised learning is used when only a small portion of the data has labels.
Deformable objects is a topic that is considered much harder than the one with rigid objects due to its dimensionality and for that reason it hasn't been studied as much until recently. Thus the lack of labelled data in this case is an issue that needs to be addressed.

Objective
The objective of this project is to apply the above two techniques on simulated garment manipulations and, in the case of Semi-supervised learning, estimate the amount of labeled data which is required to acquire comparable results to supervised learning.

Methodology
The steps to achieve the above objectives are:
Data generation of simulated garment deformation. The simulator is already provided although improvements are always welcome.
Application of state of the art methods(SimCLR,FixMatch etc) on state of the art models(CNNs and Vision Transformers) on the generated data.
This project can also be extended by investigating how the topological priors of the garment can be used to augment the two methods which can also lead to a publication in which the students will be named as first authors.

Prerequisites
Deep Learning expertise on computer vision, Python, Pytorch or Tensoflow(ideally Pytorch)

The project is intended for two students who will focus on state of the art deep learning methods. They should both be extremely motivated. One will mainly work on Self-supervised learning and the other on Semi-supervised learning. However their work will intersect on the simulator level and further on the deep learning framework.