We also address the problem of action retrieval: Given a single example video clip
containing an action of interest, the task is to retrieve all matching instances of that
action from an unstructured video collection. The strength of our bottom-up matching
approach is that it can compute a similarity measure between activities without learning.
It can therefore be used in situations where the space of possible activities is very large
and difficult to define a priori and when it is difficult to find an extensive amount of
training examples across different views.