Video segmentation using spectral clustering on superpixels

A spectral clustering based video object segmentation technique is proposed in this work. A foreground separation model is introduced which uses thresholding by different features to produce an initial labeling for each frame of the input sequence. We use a combination of color, optical flow, spatial-coordinates, spatiotemporal saliency and the initial foreground labeling to construct an interframe graph showing the relationship between superpixels of the entire video. The graph is solved using spectral clustering to obtain the final segmentation results. We compare our segmentation maps against state-of-the-art techniques and experimental results show that our solution is comparable to them.