eTRIMS

project results:

Online Boosting for Scene Interpretation

Jan Šochman
Center for Machine Perception
Czech Technical University in Prague
May, 2008

A novel online boosting algorithm, online WaldBoost, has been developed which reduces human intervention in the online boosting classifier design. Besides incrementality of learning, the proposed algorithm is able to automatically adapt (i) the classifier complexity, and (ii) the evaluation speed to the problem difficulty.

The algorithm is planned to be used for building an incremental appearance model for terminal symbols (e.g. windows) in the grammatical parsing algorithm. There are several advantages of such incremental approach: (i) various object types can be modelled without manually desinging their appearance models, (ii) the model is trained for specific local appearance variations and thus can be simpler than the general one, and (iii) the evaluation speed is optimised for the current problem.

Similar verification of element appearance can be found in the SCENIC system as well when the higher-level interpretation hypotheses are verified back in the image. This way aggregates, like window-door-balcony or a new type of windows can be trained and found in the image using our incremental approach. Another application could be incremental improvement of existing classifiers trained offline.







Row expansion step of the grammatical parsing algorithm combined with incremental learning. Left column shows the confidence of the incrementally updated classifier over the row positions (blue) together with smoothed confidence (red). Found window candidate position is marked by a circle in the confidence plot and by a red rectangle in the image. Note that already selected window positions and their neighbourhood are suppressed for further search.

References

[1] Jan Šochman and Radim Šára. Concepts and Modules for Autonomous Learning. eTRIMS deliverable D3.5. May 2008.

[2] Helmut Grabner, Jan Šochman, Horst Bischof, Jiří Matas. Training Sequential On-line Boosting Classifier for Visual Tracking. In review process.