Home People Research Publications Demos
         
News Jobs Prospective
Students
About Internal

TangentBoost

The design of robust classifiers, which can contend with the noisy and outlier ridden datasets typical of computer vision, is studied. It is argued that such robustness requires loss functions that penalize both large positive and negative margins. The probability elicitation view of classifier design is adopted, and a set of necessary conditions for the design of such losses is identified. These conditions are used to derive a novel robust Bayes-consistent loss, denoted Tangent loss, and an associated boosting algorithm, denoted TangentBoost. Experiments with data from the computer vision problems of scene classification, object tracking, and multiple instance learning show that TangentBoost consistently outperforms previous boosting algorithms.

Experimental Results:
Tracking

The Discriminant Saliency Tracker (DST) of [Mahadevan and Vasconcelos CVPR-09] maps frames to the feature space where the target is salient compared to the background. TangentBoost is used to combine the saliency maps in a discriminant manner.

The red tracker uses the robust Tangent loss and the blue tracker uses the Exp loss.
Tracking Demo 1
Tracking Demo 2


    

Scene Classification

We use the semantic space representation of [N. Rasiwasia and N. Vasconcelos CVPR-09] and Simply replacing the SVM with the robust TangentBoosted classifier. This results in a ~4% improvement in accuracy.

Most improvement is seen in classes with the most class confusion/outliers such as:

highway, street, tall building

mountain, forest, coast, open country

living room, bedroom, kitchen.


    


Publications: On the Design of Robust Classifiers for Computer Vision.
Hamed Masnadi-Shirazi, Nuno Vasconcelos and Vijay Mahadevan.
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2010.
Oral presentation (acceptance rate 5%)
CVPR Oral presentation video
[pdf]

Contact: Nuno Vasconcelos, Hamed Masnadi-Shirazi

 



Copyright @ 2007 www.svcl.ucsd.edu