PhD Thesis Semi-Supervised Ensemble Methods for Computer Vision

38 Chapter 3. Overview of Semi-Supervised Learning
3.8 Computer Vision Applications
In computer vision, the probably most frequently applied semi-supervised learning algorithm
is co-training. For example, Levin et al. [Levin et al., 2003] used co-training to train
a car detector. They start with a small number of hand labeled samples and generate additional
labeled examples by applying co-training of two boosted off-line classifiers, where
one uses gray-value images and the other is trained from background subtracted images,
respectively. Moreover, Javed et al. [Javed et al., 2005] applied an arbitrary number of
classifiers and extended the method to on-line learning. In particular, they first generate a
seed model by off-line boosting, which is improved later on by on-line boosting. If multiple
disjoint views exist, co-training can also be applied for tracking, e.g., [Tang et al.,
2007, Yu et al., 2008, Liu et al., 2009]. There also exist several approaches based on deep
neural networks in order to improve the visual recognition performance using unlabeled
data, e.g., [Yu et al., 2008, Mobahi and Collobert, 2009]. Recently, Fergus et al. [Fergus
et al., 2009] presented a semi-supervised framework that is able to learn object classifiers
from 80 million images. In particular, they propose a graph-based method that scales
linear with the number of samples and thus allows for large-scale usage. Guillaumin et
al. [Guillaumin and Schmid, 2010] proposed a multimodal SSL approach used for image
categorization, where the main idea is to additionally to the visual information, also
exploit other sources of information such as text, which is surrounding images on web
pages. Socher and Fei-Fei [Socher and Fei-Fei, 2010] applied Semi-supervised learning
to image-segmentation.
3.9 SSL from weakly related data
As has been shown above, there exist a large amount of methods and algorithms for the
semi-supervised learning problem. The main differences between these approaches are
often only based on their assumptions which they are imposing over the unlabeled data
(e.g., manifold assumption or large margin assumption, etc.) and on which supervised
learning method they are based, such as SVMs or boosting. Yet, one assumption that
most of them have in common is that the underlying marginal data distribution P (X , Y)
is i.i.d., which means they draw samples from data of which they assume it is identical
and independently distributed. However, in practice, unlabeled data does not necessarily
come from the same distribution as the labeled data.
Another problem that occurs in practice but is ignored by most approaches is the fact
that although unlabeled data is usually easy to obtain, unlabeled data which consists of
sufficient amounts of target class samples is not. For instance, consider the problem of
training a visual object detector for alpacas (see Figure 3.3a). For this task, it is difficult