Learning for Cognitive Vision
The main focus of this project part can be
summarized as an attempt to create more autonomous learning algorithms for
vision systems. We want to decrease the amount of human intervention which is
necessary for successful learning: selection and labelling of training images,
segmentation by hand, annotation of videos, etc. The motivation to focus
research into this direction is twofold. First, human intervention is expensive
and large scale vision systems seem feasible only if the need for such
intervention can be reduced. Second, learning in biological cognitive systems
seems to be rather autonomous with only very weak external supervision or
feedback. Several interrelated research tasks will investigate approaches
towards such more autonomous cognitive vision systems:
-
The conservative learning
framework is a bootstrap approach to learning where the performance of
different but communicating learning algorithms is improved through
co-training with minimal human intervention.
-
Complementing conservative
learning, active learning requests labelling only for the most informative
examples. We will extend our boosting algorithm for object categorization to
active learning.
-
As a main tool for conservative
learning and active learning we will research online boosting.
-
As a more technical goal we want
to reduce the amount of manual tuning for vision algorithms and instead use
learning to improve the performance of vision components.
-
Finally we will tackle the very
ambitious goal to aggregate visual primitives into combined primitives to
obtain short descriptions of scenes and objects.