Tracking for Learning

Description

Similar to Conservative Learning the main idea is to minimize the labeling effort when on-line learning an object detector. But for learning a general detector (e.g., desktop objects, faces, etc.) foreground objects can not be segmented by using background subtraction only. Thus, for this project a framework was developed for on-line learning (using incremental PCA) an object detector for "any objects" direclty from video data. Therefore an MSER based tracker (see Video 2-3) is applied for robustly extraction of training data (see Video 4). As the tracker is initilaized automatically (see Video 1) we have a full automatic on-line learning framework. Single frame detection results for different hand held objects (see Video 5-8) and faces (see Video 9) can be found in the video section.

Learning Framework

Change Detection

MSER Tracker

Videos

1. Automatically initialization of the tracker (414kB)
2. Tracking box showing MSER (538kB)
3. Tracking coke showing MSER (2.4MB)
4. MSER Tracker (left) vs. Color Tracker (right) (563kB)
5. Result - learned single frame detector: mobile phone (664kB)
6. Result - learned single frame detector: devil (704kB)
7. Result - learned single frame detector: cup (2.1MB)
8. Result - learned single frame detector: octopus (750kB)
9. Result - learned single frame detector: faces (2.4MB)

Persons Involved

• Peter M. Roth
• Michael Donoser
• Horst Bischof

  

Related Publications

[1] On-line Learning of Unknown Hand Held Objects via Tracking

[2] Tracking for Learning an Object Representation from Unlabeled Data

[3] Efficient Maximally Stable Extremal Region (MSER) Tracking