The trad itional image segmentation algorithms can be roughly divided into two types,i.e.,unsupervised bottom-up segmentation algor ithms and supervised segmentation algorithms with interactions.For complex scenes or complicated targets,the former usually fails to work well due to the lack of prior knowledge.And the latter can achieve satisfactory segmentations with users’ in teractions,but it greatly increases the burden of the users.Recently,co-segmentation as a weak-supervised algorithm has got mo re and more attention.In this paper,we propose a heat co-diffusion based image co-segmentation algorithm.We first establish a 3D co-diffusion network between images by connecting conduction edge with similar objects.Then,the image co-segmentation is converted into how to get the maximal marginal gain in the conducting network.It is proved that the problem could be solved by sub-modular theory.Compared with several state of the art co-segmentation methods,the experimental results of t he proposed method show good performance on benchmark datasets.