Adaptive Features for MRF based Cosegmentation
Huong Ninh, Guee-Sang Lee · 2016
Cosegmentation is a problem of specifying common objects from a set of images. The challenge of cosegmentation is addressed in substantial variation of the object's appearance such as shape, color, or scale. In this paper, we propose superpixel-level features which combine both color histogram and dense Sift features, is named SP-dcoSift feature. The cosegmentation problem is solved using an optimization framework. Firstly, every image is labeled into foreground / background regions by the graph based Markov Random Field (MRF) method. Secondly, image segments are matched across images by minimizing the square distances. Through each loop, the foreground and background regions are updated. We evaluate the performance with MSRC-v2 dataset which contain challenging images for object cosegmentation. To present the robustness of our approach, we compare the accuracy measurement with previous state of the art methods.