Abstract: Pixel-level object segmentation in videos is a research hotspot in the field of computer vision. Unsupervised video segmentation without user annotation imposes higher requirements on segmentation algorithms. In recent years, the modeling methods based on inter-frame motion information are often used, that is, the motion information such as optical flow is used to predict the target contour, and the model is built based on features such as color for segmentation. Concerning the problems such as confusion of foreground and background and the rough edges caused by these methods, this paper proposes a video object segmentation method that combines fully convolutional neural network features. Firstly, the contour of the salient object in the video sequence is predicted through fully convolutional network and modified combining with motion saliency label obtained by optical flow. Then a time-space diagram model is established, the final segmentation result is obtained by using the graph cut method. The proposed method is evaluated on SegTrack v2 and DAVIS general datasets. The results show that the proposed method has better segmentation performance than the method based on inter-frame motion information.

Key words: video segmentation, object segmentation, deep feature, unsupervised, fully convolutional network