Spatiotemporal semantic video segmentation