Dynamic Scene Segmentation through Object Hypotheses Ranking

A novel framework for highly dynamic scene segmentation through foreground hypothesis is developed here. This framework enables robust foreground segmentation by ranking object hypothesis over spatial space to achieve consistent object candidates and binary segmentation of a video sequence. Inside object candidates derived from spatial features in each frame are first estimated. This is followed by ranking the object candidates over a specific hypothesis space so as to yield consistent and dense object proposals. An efficient higher-order graph-cut method is adapted to optimize a Markov Random Field (MRF) model, which is instantiated by the estimated foreground hypothesis with highest score. We demonstrate the performance of our approach through experimental evaluation on a typical dynamic scene benchmark from Freiburg-Berkeley Motion Segmentation Dataset. Compared with a state-of-the-art algorithm, our method achieves improved and robust segmentation performance when dealing with highly dynamic image sequences. The segmentation accuracy of the proposed method improved by 10.19% and 92.66% pixels are correctly classified.