Research group: SlipGuru, DISI, Università di Genova (Italy)
Research interests: Statistical learning, Unsupervised classification, Feature selection methods, Face detection and recognition, Applications of computer vision and statistical learning methods
What I'm doing
2005. I spent this year acquiring the necessary background knowledge on representing images in terms of features and studying and comparing some different methods to approach the feature selection task.
I implemented the computation of integral image and rectangle features proposed by Viola and Jones in 2001.
I developed a feature selection method based on a simple measurement on positives and negatives distributions.
Finally I have compared three different methods for feature selection: "Landweber", "Adaboost" and "Distribution based"; my experiments shown that Landweber algorithm, based on statistical learning theory, seems promising and requires a deeper study.
In the last part of my first year I started a collaboration with a company working on computer vision commercial products, the idea is to develop a real face recognition application working under uncontrolled conditions. I'm already working on a new framework for video capture, image processing and video output.
PhD thesis proposal: pdf, slides
2006. At the beginning of the second year of my PhD I made some experiments on color-based face detection, doing pixel color analysis on YUV space to detect human skin, and evaluating texture on skin regions, to discard false positives detected by color. The result we obtained was quite good in controlled conditions, and a prototype of our system was presented at the exhibit "Sicurezza 2006".
I also tested extensively a method for feature selection in a feature-based face detection task, based on the Thresholded Landweber algorithm proposed in a very different scenario. The method proved to be quite robust and comparable to Viola and Jones algorithm for feature selection.
I also tested directly the two different approaches to the same dataset, which I built using a system installed in a setting useful to test our methods in a real-world scenario.
I explored the importance of feature correlations in this particular task, and I did some experiments in building graphs of non-trivially correlated features; those features gives good results in terms of classification performance.
PhD thesis progress report: pdf, slides