Object detection using a shape codebook

This paper presents a method for detecting categories of objects in real-world images. Given training images of an object category, our goal is to recognize and localize instances of those objects in a candidate image. The main contribution of this work is a novel structure of the shape codebook for object detection. A shape codebook entry consists of two components: a shape codeword and a group of associated vectors that specify the object centroids. Like their counterpart in language, the shape codewords are simple and generic such that they can be easily extracted from most object categories. The associated vectors store the geometrical relationships between the shape codewords, which specify the characteristics of a particular object category. Thus they can be considered as the "grammar" of the shape codebook. In this paper, we use Triple-Adjacent-Segments (TAS) extracted from image edges as the shape codewords. Object detection is performed in a probabilistic voting framework. Experimental results on public datasets show performance similiar to the state-of-the-art, yet our method has significantly lower complexity and requires considerably less supervision in the training (We only need bounding boxes for a few training samples, do not need figure/ground segmentation and do not need a validation dataset).