Our current implementation of an object detector detects objects using an Xbox Kinect. It is given a model of the desired object, and then it samples areas of the scene that could correspond to the model. It ranks its hypotheses based on how well they match the model’s image features, such as surface normals, edges, and principal curvatures, and then it returns a point estimate of where it believes the object to be in the scene. For my SuperUROP project, I propose to modify the object detector in order to improve its efficiency and reliability by 1) outputting a probability distribution and 2) inputting a prior distribution.

I worked in the CSAIL LIS lab run by Leslie and Tomas as a UROP under Jared Glover my junior year, and over this past IAP, we wrote a research paper that was accepted into the 2014 AAAI Conference on Artificial Intelligence! Also, I have already worked a little with the object detector mentioned above.