Combinatorial and statistical methods for part selection for object recognition

Abstract

In object recognition tasks, where images are represented as constellations of image patches, often many patches correspond to the cluttered background. In this paper, we present a two-stage method for selecting the image patches which characterize the target object class and are capable of discriminating between the positive images containing the target objects and the complementary negative images. The first stage uses a combinatorial optimization formulation on a weighted multipartite graph. The following stage is a statistical method for selecting discriminative patches from the positive images. Another contribution of this paper is the part-based probabilistic method for object recognition, which uses a common reference frame instead of reference patch to avoid possible occlusion problems. We also explore different feature representation using principal component analysis (PCA) and 2D PCA. The experiment demonstrates our approach has outperformed most of the other known methods on a popular benchmark dataset while approaching the best known results.

Keywords:

• Class recognition
• Computer vision
• Feature selection
• Object detection
• Pattern representation and modelling

Notes

^†It must be remarked that this model extends to modelling multiple object classes directly; however, since our problem consists of only one class, we have P(O _j |A _i )=1.

^† http://www.vision.caltech.edu/html-files/archive.html