Exjobbsförslag från företag

Detta är ett uppsatsförslag hämtat från Nationella Exjobb-poolen. Klicka här för att komma tillbaka till samtliga exjobbsförslag.

Förslaget inkom 2011-01-12

Machine learning under high intra-class variation with applications to visual object detection

In machine learning, problems are often formulated as nonlinear classi?cation of high dimensional data. Design of such classifiers is challenging and over the recent decades Boosting learning methods [Y. Freund and R. Schapire. A short introduction to boosting. 1999] have proven to be effective. Given weak classifiers (performing slightly better than chance), AdaBoost is one such method that provably can achieve arbitrarily good generalization bound. Perhaps the most demonstrating paper in applications of AdaBoost for detection discriminancy, is still the famous Viola-Jones paper [P. Viola and M. Jones. Robust Real-time Object Detection. 2001]. Using the highly ef?cient concept of integral images, they showed that AdaBoost could create the ideal building-blocks of a cascade of strong classifiers, for the task of visual object detection (ie. face detection or pedestrian detection). Rasolzadeh et al. showed [B. Rasolzadeh, L. Petersson and N. Pettersson. Response Binning: Improved Weak Classifiers for Boosting. 2006] that there is room for much improvement on the weak classifiers of this algorithm by introducing the concept of multi-thresholding and response binning.

However, one major difficulty that still remains is to find an efficient way of dealing with the high intra-class variation that typically arise in visual object detection applications. We have recently explored an often glossed over aspect of the Boosting framework: the analysis of co-dependency of weak classifiers for creating what we call Gated Classifiers. This type of classifiers seem to improve the ability of the learner to handle intra-class variation. The thesis will involve comparing different strategies (including gated classifiers) for handling target classes with high intra-class variation. The results should be related to concepts in learning theory and provide guidelines for practitioners in the field. The project will be done in collaboration with the Computer Vision and Active Perception Lab (CVAP) at KTH and OculusAI AB (www.oculusai.com), a young IT company in the computer vision industry.

Objective & requirements
The student is expected to perform a series of practical experiments using a large industrial dataset, provided by OculusAI. The student is expected to have a good background in machine learning and its theoretical foundation. Also, since the applications and experiments are on visual data (ie. images), a background in computer vision or image processing is desirable.

Duration & compensation
The project is intended to start ASAP, and continue for 6 months. According to a mutual agreement on terms and deliverables (that will be set upon starting the project) there is a possibility for financial compensation after completed thesis.

Supervision will be done by the two PhD students behind the Gated Classifier method, Oscar Danielsson (CVAP/KTH) and Babak Rasolzadeh (CVAP/KTH & OculusAI). The student will spend time both at the CVAP institution at KTH as well as the offices of OculusAI in Stockholm.


Informationen om uppsatsförslag är hämtad från Nationella Exjobb-poolen.