spacer
School of Computer Science The University of Adelaide Australia
Computer Science Home
Staff Only
text zoom: S | M | L

School of Computer Science
Level 4
Ingkarni Wardli Building
THE UNIVERSITY OF ADELAIDE
SA 5005
AUSTRALIA
Email

Telephone: +61 8 8313 4729
Facsimile: +61 8 8313 4366


You are here: Computer Science > users > paulp> GSLDA

Sakrapee (Paul) Paisitkriangkrai
Postdoctoral Research Fellow
The Australian Center for Visual Technologies
School of Computer Science, The University of Adelaide, Australia

Level 5, Innova 21, The University of Adelaide, Adelaide, SA 5006, Australia
Phone: +61 8 8313 0282
Email: paulp(at)cs(dot)adelaide(dot)edu(dot)au
Google Scholar: http://scholar.google.com.au/citations?hl=en&user=lNqlhO4AAAAJ
Web: http://www.cs.adelaide.edu.au/~paulp



Research Projects

Object detection using Greedy Sparse Linear Discriminant Analysis (GSLDA)

Based on our observations, AdaBoost is sub-optimal for training visual object detector since it operates under the assumption that the number of positive and negative samples are equal. In other words, it ignores the fact that the training data in object detection problem is often imbalance and highly skewed. In order to further improve the performance, we introduce a new classifier, termed Greedy Sparse Linear Discriminant Analysis (GSLDA), for its conceptual simplicity and computational efficiency. Unlike Adaboost, GSLDA takes the number of training samples in each class into consideration when solving the optimization problem. This extra information helps minimize the effect of imbalanced data sets and improves the overall classification accuracy at the same runtime cost.

One major drawback of GSLDA classifier is that decision stumps' thresholds are fixed for the entire duration of classifier training, i.e., once calculated, we do not re-train these weak classifiers. Having a fixed threshold value could result in a suboptimal weak classifier. We propose a new technique, termed Boosted Greedy Sparse Linear Discriminant Analysis (BGSLD), to efficiently train a weak classifier. BGSLDA exploits the sample re-weighting property of boosting to update weak classifiers' thresholds and the class-separability criterion of GSLDA to train a strong classifier.

 


Homepage Resume Publications Projects