Computing object-based saliency via locality-constrained linear coding and conditional random fields