Using visual segments and spatial layouts for unsupervised object co-segmentation

Given a set of images containing the instances of the same object class, the proposed method not only partitions every image into object and background, but also parses every object into several visual segments. Unlike the semantic parts based on high-level concepts, the visual segments prefer focusing on the low-level visual features which are easier to find and match from one image to other image by simple similarity measurement. Towards this goal, an iterative process is performed in the image set, including the appearance models learning and the energy function minimizing. The initial appearance models are learned from the image set according to the saliency measurement and the objectness measurement, including the segment models and the background model. Specifically, a novel inter-image constraint is exploited in the energy function by using the layout-based similarity measurement. By experiments on a variety of image datasets, the proposed approach efficiently segments and parses the object instances with varying appearance and shape, under challenging environmental conditions.