Speeding up dominant object video segmentation

In this paper, we propose an enhancement to one of the most accurate techniques in the literature for dominant object video segmentation to achieve almost the same segmentation accuracy in less computation time. We propose the speed up of the computation by downsampling the input video frames based on the speed of the objects in the video. The average of median value of the optical flow of the video frames is used to decide the downsampling percentage to apply. Three cases are identified, downsampling by 67% (skip two frames every three frames for slow moving objects), 50% (skip frame every two frames for moderate speed moving object), and 0% (no frame skipping for fast moving objects). We generate object proposals for selected frames and use Layered DAG based approach to select the primary object using dynamic programming. The primary objects are then estimated in the unselected frames using linear interpolation of the optical flow. GMM and MRF are also used to ensure an accurate pixel-level video segmentation. The proposed approach is tested using SeqTrack [15] datasets and achieves a speed up of (50%-63%) without much compromising of the segmentation accuracy.