Enhancing boundary for video object segmentation

Video object segmentation aims to separate objects from background in successive video sequence accurately. It is a challenging task as the huge variance in object regions and similarity between object and background. Among previous methods, inner region of an object can be easily separated from background while the region around object boundary is often classified improperly. To address this problem, a novel video object segmentation method is proposed to enhance the object boundary by integrating video supervoxel into Convolutional Neural Network (CNN) model. Supervoxel is exploited in our method for its ability of preserving spatial details. The proposed method can be divided into four steps: 1) convolutional feature of video is extracted with CNN model; 2) supervoxel feature is constructed through averaging the convolutional features within each supervoxel to preserve spatial details of video; 3) the supervoxel feature and original convolutional feature are fused to construct video representation; 4) a softmax classifier is trained based on video representation to classify each pixel in video. The proposed method is evaluated both on DAVIS and Youtube-Objects datasets. Experimental results show that by considering supervoxel with spatial details, the proposed method can achieve impressive performance for video object segmentation through enhancing object boundary.