| Authors |
Armin Berger, Roth Peter M., Leistner Christian, Bischof Horst |
| Appeared in |
Proc. Workshop of the Austrian Association for Pattern Recognition |
| Date |
2010 |
| Abstract |
Recently, multi-camera networks have proven to be a valuable tool for learning object detectors. By
exploiting geometry information given by homographies between different cameras in a co-training
framework very valuable training samples can be acquired. However, for a growing number of cameras
the existing methods become infeasible for two reasons: (a) The number of required camera-to-
camera homographies increases dramatically and (b) the information fusion becomes more and
more complicated. To overcome these drawbacks, we propose a centralized approach to fuse the information
from different cameras during co-training. In particular, all detections are projected onto
the common top view and are merged using a mean shift approach. To finally generate the updates,
the thus obtained merged detections are re-projected to the specific camera views. We demonstrate
our approach for the task of person detection, where we show that even using only a small number of
labeled training samples finally state-of-the-art detection results can be obtained. |
| Link |
PDF |
