| Authors |
Max Stricker, Roth Peter M., Bischof Horst |
| Appeared in |
In Proc. Computer Vison Winterworkshop |
| Date |
2012 |
| Abstract |
When learning classifiers for tasks such as object recognition or categorization
typically a large amount of training data is required to cope with the high
intra-class variation. Since this data is often hard to obtain, in this paper we
propose an approach to automatically generate labels for images in video
streams. In particular, we use slow feature analysis (SFA) to detect the most
dominant class changes within unlabeled video sequences by analyzing the slowest
varying signals. Thus, in contrast to existing methods building on tracking,
which assume similarity of temporally close frames, also greater changes in
appearance can be handled. To make the approach more robust, as underlying
representation we apply a Bag of Word (BoW) model of SIFTs, which is obtained
via sparse coding. These findings are confirmed in the experimental results,
where we show the class separation capability as well as classification results
for two different data sets. |
| Link |
|
