Scene Classification Using External Knowledge Source

Abstract

In this paper, we introduce a model for scene category recognition using metadata of labeled training dataset. We define a measurement of object-scene relevance and apply it to scene category classification to increase coherence of objects in classification and annotation tasks. We show how our context-based extension of supervised Latent Dirichlet Allocation (LDA) model increases recognition accuracy when feature mix is influenced by our relevancy score. We demonstrate that the proposed approach performs well on LabelMe dataset. Comparison between our purposed approach and state of art semi-supervised clustering algorithms using labeled data shows effectiveness of our approach in interpretation of scenes.