Conservative Learning

Description

The motivation for Conservative Learning was to minimize the manual effort when learning a classifier and to combine the power of a discriminative classifier with the robustness of a generative model. Thus, a simple motion detection, an on-line AdaBoost algorithm and a robust generative model are combined. To obtain the final classifer a discriminative classifier is trained on-line where the update rules are defined by the robust generative model. The on-line updating is visualized in Video 1 (green - pos. update, red - neg. update, white - detection but no update) and incremental better results are shown in Video 2. Single frame detection results for different scenarios are shown in Video 3-5.

Learning Framework

Update Rules

Results

Videos

1. Conservative Learning visualized (643kB)
2. On-line classifiers compared (515kB)
3. Final detection results - CoffeeCam (1.4MB)
4. Final detection results - Caviar (1MB)
5. Final detection results - Tunnel (2.1MB)

Persons Involved

• Peter M. Roth
• Helmut Grabner
• Danijel Skocaj (University of Ljubljana)
• Horst Bischof
• Ales Leonardis (University of Ljubljana)

Related Publications

[1] On-line conservative learning for person detection

[2] Conservative visual learning for object detection with minimal hand labeling effort

[3] On-line Learning a Person Model from Video Data

[4] On-line Boosting and Vision