Spatiotemporal-sensitive video understanding