Mustafa Özuysal

I worked as a doctoral student under the supervision of Dr. Vincent Lepetit (co-advisor) and Prof. Pascal Fua. My research interests include object tracking-by-detection, camera egomotion estimation and augmented reality. My thesis work concentrates on learning features from image and video sequences for fast and reliable object detection.

My private thesis defense took place on 01.06.2010 and I received my Ph.D. degree after the public defense on 06.07.2010 ...

A Naive Bayes Multiview Pose Estimator for Object Categories
The main challenge of multiview object detection is the variation of image features as the object is imaged under different viewpoints. We show that using a pose estimator to select a pose-tuned classifier, it is possible to improve localization results and to provide object pose information at the same time as detection.

A Semi-Naive Bayesian Classifier (Ferns) for Fast Patch Classification
We show that it is possible to train a Semi-Naive Bayesian Classifier to recognize keypoints without the need for intensive preprocessing stages like fine-scale selection, dominant orientation and affine parameter estimation. We also demonstrate that this approach is superior to previous methods that use weighted averaging of Randomized-Tree based classifier ensembles.

Source: Web, Licence: Creative Commons Attribution-ShareAlike 2.0

Rigid 3D Object Detection and Tracking (Feature Harvesting)
To initialize and improve the robustness of point feature based tracking systems, we have developed a method that automatically learns the appearance of the keypoints on the object to be tracked by incrementally training a Randomized Tree based classifier. The learned classifier can then be used to detect the object in a given image and camera pose computed in the reference coordinate system of the object model.