PhD Thesis Semi-Supervised Ensemble Methods for Computer Vision

5
ing algorithms. Additionally, in practice there exist many scenarios where the data arrives
sequentially and systems have to be updated without re-training the entire model. This
is, for example, the case in robot navigation or visual object tracking. Finally, in order
to learn from large amounts of data, as argued in the SSL literature, it is essential to apply
algorithms that are computationally efficient and can operate with limited amounts of
memory. On-line learning methods are faster and require less amount of memory compared
to off-line algorithms which makes them ideal for semi-supervised learning tasks.
Object Tracking Semi-supervised learning methods are mainly applied in order to increase
the accuracy of a classifier by exploiting additional huge amounts of unlabeled
data. However, SSL can also be applied on a task which, at the first glance, does not seem
to be related to learning from both labeled and unlabeled data, i.e., visual object tracking.
In fact, if we consider a tracking task, where an a priori unknown object is marked in the
first frame a powerful tracking approach is to learn a binary classifier with the marked
patch as positive sample and the surrounding patches as negatives, respectively. Then, the
tracking of the object is performed by using this classifier to re-detect the marked object
in the subsequent frames. Hence, such approaches are also called tracking-by-detection
methods. In order to make the classifiers adaptive towards rapid appearance and illumination
changes, typically the classifier updates itself on the re-detected object, according to
hand-designed update rules. As already mentioned, such approaches yield highly accurate
trackers which are also fast because the object has only to be discriminated versus its local
background and the model complexity can be kept very low. Nevertheless, one problem
that all of these methods have in common is that slight errors during the self-updating
process can easily accumulate and finally lead to failure of tracking, i.e., the object is
lost. This is also called drifting. One reason for drifting of these tracking-by-detection
approaches lies in the fact that supervised learners are abused for an in principle unsupervised
learning task. In fact, one can easily see that labeled data is only available at the
first frame when the object is marked. In all subsequent frames the task of the tracker is
to act autonomously and without getting any additional labels for the data it is observing.
Hence, the learner over time has to be able to exploit both labeled data and unlabeled data,
which is a natural semi-supervised learning problem. Following this observation, in this
thesis, we will propose novel on-line semi-supervised learning algorithms and we will
show that applying these methods to visual object tracking results in more robust tracking
results.