Shape Detection with Nearest Neighbour Contour Fragments

Shape is probably the single most important feature for object detection and much research has gone into developing deformable shape models. However, contours extracted by bottom-up edge detectors are notoriously unreliable, especially in natural images. In this paper, we present a very simple, yet powerful method for model creation, hypothesis generation, and hypothesis verification, which is competitive with much more complex methods. Let s be a descriptor in some high-dimensional space of an edge fragment representing part of an object contour. In Bayesian terms, this fragment was likely to be generated by some class c∈C if the conditional likelihood P(s|c) is greater for c than that for any other class, P(s|c′ 6= c), including the background class. P(s|c) can be estimated non-parametrically, for example by using Gaussian kernels associated with some nearby samples in the feature space. In [2] it was shown that a good approximation can be obtained by using only the nearest sample. A fragment is discriminative if it is much more likely to belong to a class c than any other class c′ 6= c. We thus define the discriminative power of a fragment by: