Toom

XNA is a framework developed by Microsoft to allow developers to design games more easily. The bright side of this framework is that it's working on all Microsoft platforms (PC, Phones, XBOX 360...). During my last year at the NUI Maynooth I had to develop two programs using XNA and VisualStudio (C#). One XNA project was to develop a small 2D game and the other was to render and move around a 3D world using Quaternions. For the second one I decided to create a very simple environment (using Google Sketchup) where you can drive a very simple remote control car (see picture below).

3D game sources are available as usual.

For the 2D game I decided do to something more complex and more interesting. This game is using the concept "destroy-n-build". You are in an environment composed of blocks and you have to mine them and to build thanks to them to reach the bottle. You can play up to 5 players on one computer; 4 players will use XBOX 360 gamepads and the last player will use the keyboard and the mouse. Another great thing is the possibility to create your own worlds and levels with a simple text editor. The picture below shows a party with 2 players in a "big" level.

The UML of this program is clear, strong and allow upgrades. Moreover the architecture can be re-used to create another 2D game. It was generated thanks to yUML and my yUML file is available here.

And here are the 2D game sources

The BCC is a fast detection technique using Haar-like features to recognize shapes in an image. This technique is used in lot of devices, e.g. cameras to detect faces. The main advantage of this technique is the speed to execute the detection over an image. The BCC is based on the neural network called rejection cascade. The aim of this cascade is to reject non-shapes with an increasing accuracy at each stump (node/stage) of the cascade. The picture below synthesizes the process.

To train a cascade like this, the user will need a database of positive and negative images. The two pictures below represent typical positive (top) and negative (bottom) images for a robot detection BCC. Of course an image of a fish could be considered as a negative image in this case but keep in mind that positive and negative images should be in a common environment.

Then the trainer will compute Haar-like features (using Haar-like masks with different size, positions and orientations) that are common between positive images and not found in negative images. Finally it will create different stumps (small neural networks) with different accuracies and order them by accuracy (less accurate to most accurate) to create the BCC. Haar-like masks are simple binary masks (see picture below), the main advantage of these masks is the possibility to apply them to an image using Fourier transform and then to speed up the process.

With a well-trained cascade (database of 1000 positives/1000 negatives, and 20 stumps) approximately 70% of non-shapes are rejected by the three first nodes. The fact that less accurate stumps are at the beginning of the cascade is the key of the speed of execution.

The main downside is the time to train the cascade: from one hour to a month depending on the database size. Another downside is that all images in the training set (database of images) have to be clear (no blurred images) because Haar-like features are detecting sharp differences between two areas in the image.

Even though new techniques were discovered after the Viola & Jones paper, the BCC is still one of the most powerful technique to detect shapes. That's why I developed a software to simplify the creation of databases of images, and that's why I used this technique to detect NAOs on a field for the RoboCup SPL.