This paper presents a promising technique to downsample for rapidly downsampling spatial dimensions in CNNs. It significantly reduces memory requirements and can be expected to not have much runtime overhead. The experiments are well done and the reviewers agree the paper should be accepted.