Abstract: An unsupervised segmentation method is proposed in this paper, which collaboratively segments multiple images depicting similar object into foreground and background. Given a set of images, a co-objectness map is computed as the hint about possible foreground locations according to both inter-image and intra-image cues. The map is used as a geometric prior, which describes the location and shape information of object. Then, by bi-segmenting the map, it gets a rough estimation of foreground/background labeling, based on which, a binary classifier of superpixel is trained. The classifier can propagate the appearance distribution（i.e. appearance prior） across the images. Finally, the geometric and appearance priors are incorporated into a conditional random field to achieve object segmentation. The method is tested on the MSRC dataset and recently introduced iCoseg dataset, and shows better results than the state-of-the-art.

Key words: object segmentation, co-segmentation, objectness, conditional random field

摘要： 提出一种协同分割算法，使包含同类目标的多幅图像相互作用，从而将目标从各自图像的背景中分离出来。首先，分别从单幅图像自身角度和多幅同类目标图像相互作用的角度出发，计算出图像中每个像素或区域属于前景或背景的似然概率，从而得到协同目标性映射图。这个映射图描述了目标的位置和几何形状信息，然后阈值化这个映射图作为图像分割真值来训练一个关于超像素的二值分类器，用训练好的分类器预测出每个超像素的前背景似然概率作为外观先验信息，与几何先验信息一并送入条件随机场模型，从而实现对图像目标的分割。在MSRC和iCoseg两个数据库上的测试结果表明该算法的分割效果优于同类方法。

关键词: 目标分割, 协同分割, 目标性, 条件随机场