Unsupervised Multi-class Cosegmentation

Cosegmentation aims to segment out similar objects from a set of images with minimum additional information. Most of the existing cosegmentation algorithms assume that the foreground objects should appear in all images of the image set. But under some conditions, if the foreground objects only appear in a few images, the segmentation results are possible to be wrong. The paper proposes a new cosegmentation algorithm which can segment and classify the foreground of different objects even if they do not appear in all images. In our work, an image is considered to contain several kinds of objects. Each object is composed of several object elements; therefore, each image can be expressed in terms of the combination of several object elements. Object elements with similar features could be grouped into one object-element cluster by using a density-clustering algorithm. Moreover, the density-clustering algorithm excludes a few object elements which do not have a sufficient number of similar object elements. During the segmentation process, we de-project the sub-object classes back to images. Observing the distribution of each sub-object classes, we select the appropriate classes as the segmented results through the selection criteria. In the work, an unsupervised multiple-object class framework is proposed, and the segmentation rate is enhanced by introducing the concept of independent object elements. A selection criterion is presented to relax the similar object constraint.