Abstract

We present a video object segmentation approach thatextends the particle filter to a region-based image repre-sentation. Image partition is considered part of the parti-cle filter measurement, which enriches the available infor-mation and leads to a re-formulation of the particle filter.The prediction step uses a co-clustering between the pre-vious image object partition and a partition of the currentone, which allows us to tackle the evolution of non-rigidstructures. Particles are defined as unions of regions in thecurrent image partition and their propagation is computedthrough a single co-clustering. The proposed technique is assessed on the SegTrack dataset, leading to satisfactory perceptual results and obtaining very competitive pixel error rates compared with the state-of-the-art methods.