Human-Centered Saliency Maps without Sacrificing Accuracy?