Independent thesis Advanced level (professional degree), 20 credits / 30 HE credits
Robust Factorized Shape Descriptors
by Victoria Matute Arribas
Imagine a robot driving a car. It has GPS signals that tells it which path to follow. However, it should adapt to the traffic conditions in every place.
It wants to go from Gotemburg to Stockholm. Traffic is chaotic that morning in Gotemburg. There are some traffic police leading traffic.The robot has learned all the traffic signs that the police use to direct the traffic.
What should do the robot internally? Most of the signs are made with hands. The robot should look for something that looks like hands independently of the orientation of these hands. It will focus on the appearance of the objects. Our descriptor should be independent to rotation but not to appearance.
Once the hands have been located, the robot should figure out what the police wants it to do. If the hand is vertical, the robot should stop. If the hand is horizontal to the left, it will have to turn left. In this case, the orientation of the hands matters. Our descriptor should be independent to appearance but not to rotation.
The robot arrives to Stockholm. It wants to go to the city center. It finds some road signals with different directions depending on the destination. Stockholm has changed significantly during the last decade and there are both old and new road signals. The shape of the arrows are different for the new and old signs. Which patterns should the robot focus now? Rotation or Appearance? Appearance is important to focus only on arrows. However, it is more important the rotation of the arrow in order to follow the right way. In this thesis, we present a factorized shape descriptor.