Video Object Segmentation In Unconstrained Videos

Video object segmentation,which is pixel-level segmentation of key targets in video,has been widely used in content-based retrieval,intelligent monitoring,video conferencing and other fields.In recent years,many excellent algorithms have been proposed,but due to a series of interferences such as non-rigid deformation,strenuous motion and target occlusion in unconstrained videos,this technology faces great difficulties and challenges.The unsupervised video object segmentation has no user annotation at all,and it puts higher requirements on the segmentation algorithm.How to design a robust,fast and accurate video object segmentation system to meet actual needs has very important research significance.The unsupervised video object segmentation algorithm is commonly used in the pretargeting and fine-segmentation architecture,the extraction of appearance and motion information,and the optimization of the graph model of videos are the research focuses in this field.The recent introduction of a full convolutional network has made the segmentation algorithm based on the full convolution network a hot research topic,and has made great progress in time and precision.This paper applies the segmentation method based on full convolutional network,combined with the original video object segmentation framework,and analyzes and improves its problems.The main contributions include:An unsupervised video object segmentation algorithm based on appearance significant network is proposed.For the problems of foreground background confusion and edge roughness that may exist in the method based on motion information initialization,the framework of FST is adopted,and three parts of optimization are carried out based on it: 1)using the appearance significant network to extract a more targeted appearance saliency map;2)use SIFT optical flow algorithm and cross-bilateral filtering to extract more accurate motion prediction labels;3)apply full-connection condition random field to further refine the edge of segmentation.Qualitative and quantitative evaluations were performed on the Seg Track v2 and DAVIS datasets.Compared with the FST,there were 6.3% and 7% accuracy improvements on the two datasets respectively.The final segmentation effect solved the problems like background confusion and rough edges well.