Pairwise Grouping Using Color

Abstract

Grouping was recognized in computer vision early on as having the potential of improving both matching and recognition. Most papers consider grouping as a segmentation problem and a hard decision is made about which pixels in the image belong to the same object. In this paper we instead focus on soft pairwise grouping, that is computing affinities between pairs of pixels that reflect how likely that pair is to belong to the same object. This fits perfectly with our recognition approach, where we consider pairwise relationships between features/pixels. Some other papers also considered soft pairwise grouping between features, but they focused more on geometry than appearance. In this paper we take a different approach and show how color could also be used for pairwise grouping. We present a simple but effective method to group pixels based on color statistics. By using only color information and no prior higher level knowledge about objects and scenes we develop an efficient classifier that can separate the pixels that belong to the same object from those that do not. In the context of segmentation where color is also used only nearby pixels are generally considered, and very simple color information is taken into account. We use global color information instead and develop an efficient algorithm that can successfully classify even pairs of pixels that are far apart.