Sections
 
You are here: Home › Members › Michael Donoser › Interest Point Tracking

Tracking of Maximally Stable Extremal Regions

You can find more details on our team homepage at Virtual Habitat:http://vh.icg.tugraz.at

The detection of interest points and local features constitutes the basis for many important computer vision tasks. For example, object recognition, stereo matching, mosaicking, robot navigation etc. rely on the detection of interest points which possess some distinguishing, highly invariant and stable properties. Such structures are often called distinguished regions (DR) and provide a compact and abstract representation of patterns in an image.

Numerous interest point detection algorithms have been proposed like Harris-Affine and Hessian Affine, edge-based detectors, intensity extrema-based detectors, salient regions or random tree based detectors. Detailed evaluations and comparisons of different interest point detectors are available. They revealed that the Maximally Stable Extremal Region (MSER) detector from Matas et al. performs best on a wide range of different test sets. MSERs denote a set of distinguished regions which are defined by an extremal property of its intensity function in the region and on its outer boundary. MSERs have all the properties required of a stable local detector. These properties make them also capable of segmenting homogenous regions within images as is shown in our work on unsupervised color segmentation (http://www.icg.tugraz.at/Members/donoser/do_unsegmentation)

If a sequence of images is available as input for interest point detection temporal information can be included to improve the overall detection quality. For example, Video Google describes an approach to object and scene retrieval based on tracked distinguished regions, where tracking and interest point detection are realized by different algorithms. Obviously, results would be improved if both detection and tracking would be based on the same principles. In conclusion, detection and stable tracking of distinguished regions through an image sequence is of high interest.

We have introduced an algorithm for detection and tracking of Maximally Stable Extremal Regions (MSERs). The scientific contribution was twofold, first, the proposed method improves the computational time of MSER detection and second, by means of tracking more stable results compared to single frame MSER detection are obtained.

We have also shown ways to integrate color and shape into the MSER Tracking framework, which allows to use it e.g. for face and hand tracking or to integrate it in an unsupervised online learning framework.

Interest Point Tracking

Publications

  • Donoser, M. and Bischof, H. (2008). Fast Non-Rigid Object Boundary Tracking. In Proceedings of British Machine Vision Conference (BMVC), Leeds, GB.
  • Donoser, M. and Bischof, H. (2008). Real Time Appearance Based Hand Tracking. In Proceedings of International Conference on Pattern Recognition (ICPR), Tampa, USA
  • Donoser, M. and Bischof, H. (2006). Efficient Maximally Stable Extremal Region (MSER) tracking. In Proceedings of Conference on Computer Vision and Pattern Recognition (CVPR), pages 553–560, New York, USA.
  • Donoser, M., Arth, C. and Bischof, H. Detecting, Tracking and Recognizing License Plates. In Proceedings of Asian Conference of Computer Vision (ACCV), accepted july 2007, Tokyo, Japan.
  • Roth, P., Donoser, M., and Bischof, H. (2006). On-line learning of unknown hand held objects via tracking. In Proceedings of International Cognitive Vision Workshop (ICVW).
  • Roth, P., Donoser, M., and Bischof, H. (2006). Tracking for learning an object representation from unlabeled data. In Proceedings of Computer Vision Winter Workshop (CVWW), pages 46–51.
Document Actions
[Powered by Plone]