A quick summary of my research interests
In general, I'm interested in computer vision: that is, the problem of teaching computers how to see. This is a huge research area, too big for any one brain to wrap itself around completely, so all us vision researchers focus on smaller problems within it. One day, we hope, our combined efforts will play a part in reaching that lofty goal. Below are a couple of the larger projects I've been involved with. For more detail, see my Publications page.
If you are a prospective Honours student
Feel free to get in touch with me to discuss any vision related projects. You can bring along your own idea, or alternatively we are always coming across smaller scale projects, that we would love to try out but are not necessarily related to the projects below.
If you are a prospective Masters/PhD student
Also feel free to email me, of course. However for me to supervise your project it will need to be more closely aligned with one of the problem domains I have spent more time with or am interested in pursuing long term in the future.
I'm interested in tracking lots of people or objects at once, in more than one camera. The typical challenge in these scenarios is that it's not easy to tell people apart visually, and so we have to combine large amounts of ambiguous information gathered over time to arrive at a result.
There are two applications of this I'm currently interested in (and funded to work on). First, we are trying to track Australian Rules Football players during a game, in order to provide a complete picture of where each player is at each moment in time. This is a particularly hard tracking problem because there are so many players (18 per team), running around on a huge field, and they tend to bunch together around the ball. But it's also happening in a fairly constrained environment, in a game with a fixed set of rules, which we can take advantage of when the visual information is poor.
Second, we are starting to look at how to make "smart buildings" actually live up to their name. At present, such buildings are able to gather vast quantities of data from sensors deployed throughout them, but not able to do much with it. There is great scope to do better, and to have buildings that actually respond to and accommodate the people and things within them, even with current sensor technology.
I have a longstanding interest in reconstructing the 3D properties of scenes from images. In particular, within ACVT we have been developing methods for combining geometric and semantic information (the "where" and the "what") to come up with a meaninful recreation of a scene based only on 2D images of it. Along the way we have reconstructed buildings based on architectural style, built a sketch based 3D modeller that requires no 3D interaction, and developed an augmented reality authoring system with a single button interface, among other things. It's an interesting area!
One of the more exciting directions in this area is the combination of recent advances in 2D object detection and recognition with 3D geometric information. There is lots of evidence that we humans use some combination of 2D and 3D when reasoning about our environment, and these algorithms are already delivering much richer reconstructions of scenes than can be obtained using just 2D or 3D information in isolation. For example, we are currently working on reconstructing plants from images, but in a way that recovers not only their shape but also counts the number of leaves and how they've grown since we last saw the plant.
Vision and graphics
I also have a long held interest in computer graphics, and am always looking for ways to usefully apply graphics techniques to vision, or vice versa. If you are interested in this area I suggest looking at recent SIGGRAPH proceedings to see what's new. I'm happy to discuss any papers that catch your interest.
What about machine learning?
Some background in machine learning is pretty much required to work in computer vision these days. And with good reason, as methods based on learning from data have been shown to consistently outperform those that are more "hand-crafted". Like my colleagues I am interested in the area, particularly in its application to vision, but I am more of a user of existing ML methods than a researcher with aspirations to develop new ones. If Machine Learning is your main interest I suggest talking to one of my excellent ACVT colleagues working in the area!