Video object segmentation using multiple random walkers with GMM restart rule

In this work, we propose a weakly supervised online video object segmentation algorithm, which accepts a bounding box as user annotation. First, we estimate the initial distributions of the foreground and the background by employing a visual saliency detector. Next, we simulate movements of double random walkers, one for the foreground and the other for the background. To this end, we introduce a novel restart rule based on Gaussian mixture models (GMMs). We update the GMMs during the random walk simulation to encourage interactions between the two random walkers. To achieve video segmentation, from the second to the last frames, we sequentially propagate the segmentation labels and the GMMs of the previous frame in order to maintain temporal consistency. Experimental results demonstrate that the proposed algorithm outperforms conventional video object segmentation algorithms.