Video Object Segmentation Via Enhancing Intra-connectivity And Consistency Of Object Regions

Video object segmentation is the task of separating foreground objects from the background in a video.This paper formulates video segmentation as a pixel labelling problem with two labels: foreground and background.Video segmentation is very important in computer vison,and it has a wide range of applications,including video summarization,pose recognition and target tracking.This paper presents a fully automatic technique for handling unconstrained video,e.g.,fast/slow moving objects,motion blur and cluttered background,in a superpixel-level coarse-to-fine segmentation framework.Main contributions are as follows: Firstly,a key technique is the novel spatial and temporal pairwise potential of enhancing intra-consistency of object.It uses hyper edge to encode the gestalt principle “proximity”,and this paper demonstrates its improvements in discriminative ability.Secondly,in the refinement segmentation stage,a similarity prior enhancing intra-connectivity of object is integrated into the segmentation framework as a novel unary potential,and it can indicate the likelihood of foreground effectively.Moreover,this unary potential can help segment object in frames where it is static.The proposed two-stage approach is thoroughly evaluated on DAVIS and SegTrack v2 datasets.We qualitatively show the segmentation results,and quantitatively compare the segmentation accuracy.Results show that our method outperforms the other state-of-the-art methods.