Categorization using a Discriminative Approach
Barbara Caputo (KTH Stockholm, Sweden)
In this talk I'll present an SVM-based strategy for categorization of
objects and actions in realistic settings. With respect to object
categorization, I will present a family of kernels which allow to use local
features as input of SVM classifiers, and a cue integration scheme which
accumulates confidence measures, obtained via large margin classifiers
like (but not only) SVMs, obtaining a more robust and effective
categorization.
With respect to action, I will show how the same approach developed
for object categorization can be used for multiple action
recognition in realistic settings; interestingly, experimental
results show that the algorithm tend to group together 'leg actions'
and 'arm actions'. The talk will conclude with a discussion on open
challenges for categorization of visual patterns, and some ongoing
research on those topics.