Learning-Based Local Visual Representation and Indexing

Chapter 2

Interest-Point Detection

Beyond Local Scale

Abstract

In this chapter, we propose a novel approach that introduces contextual cues in the detection stage, which is left unexploited in the existing literature. Our motivation is that, so far only local cues are used for detection, which ignores the contextual statistics from neighborhood points. As a result, it is hard to detect the real “interest” points at a higher scale. Furthermore, we also consider the possibility to “feedback” the supervised information to guide the detection stage. In this chapter, we introduce a context-aware semi-local (CASL) feature detector framework to achieve these goals. This framework boosts the interest-point detector from the traditional local scale ...