Feature Selection for conditional Bayesian networks

• by considering only regions of a minimum size and
• by focusing on stable regions.

, the scale-space layers are sketched, the stable regions are the empty rectangles, the merging events are the filled boxes. This hierarchical structure of regions is now the basis for a Bayesian network, so that we are able to propagate the classification results of regions to obtain the e.g. most probable interpretation result that is consistent with the class hierarchy of our ontology.

Conceptionally, we designed a conditional Bayesian networks, where we only observe one very high dimensional feature vector which combines the derived features from all detected stable regions. This leads to more investigations in feature selection in scale-space. We developed a feature selection scheme based on classification by two boosting methods, namely Adaboost and ADTboost. The drawing framework of ADTboost is used for the visualization of the Adaboost classifier as shown in next fig.


So far, we evaluated the classification with very simple weak classifiers and the feature selection on stable regions which have been detected in rectified images of the eTRIMS data base.


References

[1]	Martin Drauschke, Hanns-Florian Schuster and Wolfgang Förstner (2006). Detectibility of Buildings in Aerial Images over Scale Space. In: Wolfgang Förstner, Richard Steffen (Eds.), Symposium of ISPRS Commission III - Photogrammetric Computer Vision, pages 7-12. pdf
[2]	Martin Drauschke (2008). Feature Subset Selection with Adaboost and ADTboost. Technical Report Nr. 4, TR-IGG-P-2008-04. pdf
[3]	Martin Drauschke and Wolfgang Förstner (2008). Comparison of Adaboost and ADTboost for Feature Subset Selection. To appear at PRIS 2008. pdf
[4]	Martin Drauschke and Wolfgang Förstner (2008). Selecting Appropriate Features for Detecting Buildings and Building Parts. To appear at 21st ISPRS Congress 2008. pdf