Description

Within the challenging research topic of robotic manipulation of cloth, there is an issue that has received little attention, but which is of capital importance for realistic manipulation in the domestic environment. Most garments have a distinctive inner and outer side. While most of the research up to date assumes that a cloth item appears and is manipulated rightside out, the fact is that in household environments it is quite frequent that a garment is partial or totally in an inside out state. The first step towards handling such situations is a correct identification of which parts of the garment -if any- are inside out. In some cases colour information is enough for distinguishing the inside from the outside. Other cases will require the identification of specific features like seams, pockets, buttons, etc.

Objective

The objective of this project is to provide a computer vision system with the capability of identifying the parts of a garment that are inside out. A possible extension could include the automatic suggestion of actions for reversal.

Methodology

This project requires to address the following topics:

• Exploring the existing literature on image processing of garments with manipulation purposes. The focus should be set on feature-based identification of clothes.
• Acquisition of images of garments in different states. Construction of an annotated database of garments which are loosely extended on a surface and present parts that are inside out.
• In a first approach, a single class of garment will be assumed. Later, both the classification problem and the state identification problem will be dealt with simultaneously, for a limited set of garment types.

Possible extensions:

• Some actions on the cloth may enhance the perception process or resolve ambiguities. Such actions could be considered, either for general application or targeted at specific situations.
• Characterize (robot) manipulation actions for disentangling the cloth item.

The candidate will have access to the robots and cameras of our Perception and Manipulation Lab.

Pre-requisites

This project is ideal for master students with

• Some notions on Computer Vision.
• Good programming skills in C/C++