Boosting scene understanding by hierarchical pachinko allocation

Abstract

Scene understanding is a popular research direction. In this area, many attempts focus on the problem of naming objects in the complex natural scene, and visual semantic integration model (VSIM) is the representative. This model consists of two parts: semantic level and visual level. In the first level, it uses a four-level pachinko allocation model (PAM) to capture the semantics behind images. However, this four-level PAM is inflexible and lacks of considerations of common subtopics that represent the background semantics. To address these problems, we use hierarchical PAM (hPAM) to replace PAM. Since hPAM is flexible, we investigate two variations of hPAM to boost VSIM in this paper. We derive the Gibbs sampler to learn the proposed models. Empirical results validate that our works can obtain better performance than the state-of-the-art algorithms.