Semantic Context for Nonparametric Scene Parsing and Scene Classification (Author's Manuscript)
George Mason University Fairfax United States
Pagination or Media Count:
Our work focuses on different aspects of image representations as related to a variety of scene understanding tasks. We are interested in simple patch based representations as basic primitives and the role of semantic context as provided by different datasets. In our work, we have pursued a nonparametric approach for semantic parsing which uses small patches and simple gradient, color and location features. We demonstrate the value of relevance of different features channels by learning a locally adaptive distance metric and the effect of feedback in terms of semantic context, which greatly improves the performance, achieving state of the art results on different semantic parsing datasets. Here we report on an additional utility of the proposed representation for scene categorization on a subset of the scene attributes dataset introduced in.