Hierarchical Unsupervised Object Segmentation with Manifold Regularization
Zhiming Qian, Ping Zhong, Runsheng Wang · 2014
In this paper, we address the problem of object segmentation in an unsupervised way that performs image segmentation without annotated training images. To this end, we integrate low-level visual similarities with high-level semantic correlations by manifold regularization for detecting meaningful object segments. Low-level visual similarities are measured by a linear distance metric. And high-level semantic correlations are implicitly approximated from visual representations among different objects with the assumption that visual structures within an object are limited. Moreover, a hierarchical graph cut algorithm is developed for multi-class object segmentation. Finally, experimental results show a promising performance of the proposed approach on natural images.