R. Lakmper, L.J. Latecki, V. Megalooikonomou, Q. Wang, and X. Wang (USA)
: visual parts, shape similarity, machine learning,shape representation, object recognition.
In this paper we present a novel approach to learn visual parts of objects. The learned set of visual parts is optimized to yield an optimal performance for recognition of new objects. We divide the learning strategy into two main steps that follow our proposed, cognitively motivated principles of learning descriptive and distinctive parts of objects. We first extract the most dissimilar representative parts from all possible parts of all objects in the same class. We then select the most discriminative subset of this set with respect to the classification of shapes into classes. In order to computationally evaluate our approach, we developed a shape similarity measure that is able to compare parts of objects. The obtained measure yields intuitive results for significantly distorted or occluded parts even if parts are given at different scales.
Important Links:
Go Back
