Superpixels for Video Content Using a Contour-Based EM Optimization

A wide variety of computer vision applications rely on superpixel or supervoxel algorithms as a preprocessing step. This underlines the overall importance that these algorithms have gained in the recent years. However, most methods show a lack of temporal consistency or fail in producing temporally stable segmentations. In this paper, we propose a novel, contour-based approach that generates temporally consistent superpixels for video content. It can be expressed in an expectationmaximization framework and utilizes an efficient label propagation built on backward optical flow in order to encourage the preservation of superpixel shapes and their spatial constellation over time. Using established benchmark suites, we show the superior performance of our approach compared to state of the art supervoxel and superpixel algorithms for video content.