Encoding the Models with Object-Aware Feature Basis: A New Analytical Tool for Graphic Applications

Abstract

Feature space analysis is always the most central problem in all kinds of graphic applications, and the acquirement of different kinds of basis for feature space has never been stopped. In this paper, we propose a novel way to analyze the feature space by factorizing it into visually reasonable and physically meaningful basis and corresponding encoders (coefficients). Non-negative matrix factorization (NMF) has previously been shown to be powerful in information retrieval, computer vision and pattern recognition for its physically soundable and additive fashion. By transferring the factorization idea onto tasks of graphic applications, in this paper, we propose a framework for generating new feature basis and encoders for further analysis, which helps empower the downstream graphic applications, including analysis on one single model and joint analysis on a couple of models. Instead of factorizing the matrix composed of images or graphic elements/objects, we propose to apply sparseand-constrained NMF (SAC-NMF) to the feature space that is more general and extendable. And by designing various feature descriptors, we get the base functions for the feature space to enable the analysis of one single model and co-analysis of a list of models. Through the extensive experiments, our analytical framework has exhibited many attractive advantages such as being object-aware, robust, discriminative, extendable, etc.