SALISA: Saliency-based Input Sampling for Efficient Video Object Detection