Object cosegmentation by similarity propagation with saliency information and objectness frequency map
Huang Lehu, Rui Gan, Gang Zeng · 2016
Image cosegmentation is a challenge problem aimed at extracting common foreground objects from multiple similar images. To cope with greater variation of foreground appearance among images which object-cosegmentation framework can not deal with, we proposed a hybrid cosegmentation method based on object-cosegmentation framework for inter-image information and superpixel similarity propagation for intra-image information. We came up with an objectness frequency map (OFM) to measure probability of superpixels belonging to foreground and background. For object-cosegmentation part, we added saliency information as a prior cue to help refine initial common foreground objects in single image. For superpixel similarity propagation, saliency information and OFM are taken into consideration to make propagation more efficient. Common foreground objects can be segmented quickly by object-cosegmentation method and refined by superpixel similarity propogation in single image. The experimental results show that the proposed method can segment common foreground objects with lower error rates compared with previous methods.