Fast filtering-based temporal saliency detection using Minimum Barrier Distance

In the video salient object detection domain, the optical flow technique is widely used to extract temporal saliency. The big issue with this approach is that the global motion, when it exists, may be wrongly detected as a salient object. To cope with this problem, we propose to filter out the global motion by using the boundary connectivity cue. Firstly, an edge-aware filter called the guided filter is introduced to preprocess the color optical flow map for enhancing object edges. Then the pixel's connectivity to the boundary is achieved by using the minimum barrier distance in the filtered optical flow map, which leads to the effective removal of the global motion. The proposed method is assessed on the popular Seg-Track v2 and Fukuchi datasets and then compared to state-of-the-art methods. The experimental results show that the proposed approach outperforms the existing related methods.