A New Multiple Group Cosegmentation Model by Proposal Selection Strategy

文献类型： 外文期刊

第一作者： Zhu, Yin

作者： Zhu, Yin;Meng, Fanman;Xiong, Jian;Gui, Guan

作者机构：

关键词： cosegmentation;multiple group cosegmentation;EM algorithm;Markov random field

期刊名称：IEICE TRANSACTIONS ON FUNDAMENTALS OF ELECTRONICS COMMUNICATIONS AND COMPUTER SCIENCES （ 影响因子：0.338； 五年影响因子：0.324 ）

ISSN： 1745-1337

年卷期： 2017 年 E100A 卷 6 期

页码：

收录情况： SCI

摘要： Multiple image group cosegmentation (MGC) aims at segmenting common object from multiple group of images, which is a new cosegmentation research topic. The existing MGC methods formulate MGC as label assignment problem (Markov Random Field framework), which is observed to be sensitive to parameter setting. Meanwhile, it is also observed that large object variations and complicated backgrounds dramatically decrease the existing MGC performance. To this end, we propose a new object proposal based MGC model, with the aim of avoiding tedious parameter setting, and improving MGC performance. Our main idea is to formulate MGC as new region proposal selection task. A new energy function in term of proposal is proposed. Two aspects such as the foreground consistency within each single image group, and the group consistency among image groups are considered. The energy minimization method is designed in EM framework. Two steps such as the loop belief propagation and foreground propagation are iteratively implemented for the minimization. We verify our method on ICoseg dataset. Six existing cosegmentation methods are used for the comparison. The experimental results demonstrate that the proposed method can not only improve MGC performance in terms of larger IOU values, but is also robust to the parameter setting.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序