Exploring Latent Class Structures in Classification-By-Components Networks
Exploring Latent Class Structures in Classification-By-Components Networks
Lars Holdijk
Abstract
Abstract
A Classification-by-Components network (CBC) operates under the assumption that every input image can be classified based on its decomposition into a set of components. An important characteristic of these components is that they can be used in the decomposition of images from different classes. The components are class independent. In this work, we discuss the latent class structure encoded in the sharing of components between classes. We propose to visualize this structure using a Shared Component Graph (SCG). Consecutively we discuss the insight into the decision making process of a CBC the visualization can provide.
(*: denotes equal contribution)