ShuffleNet is an extremely efficient convolutional neural network (CNN) designed for mobile devices with limited computing power, such as 10-150 MFLOPs. The architecture introduces two novel operations: pointwise group convolution and channel shuffle, which significantly reduce computation cost while maintaining accuracy. Experiments on ImageNet classification and MS COCO object detection demonstrate that ShuffleNet outperforms other structures, achieving a lower top-1 error (7.8% absolute) than MobileNet on ImageNet at 40 MFLOPs. On an ARM-based mobile device, ShuffleNet achieves a ~13× actual speedup over AlexNet while maintaining comparable accuracy. The paper also explores the impact of group numbers and channel shuffle on performance, showing that larger group numbers and channel shuffle improve accuracy, especially for smaller networks. Additionally, ShuffleNet is evaluated on the MS COCO object detection task, showing superior performance compared to MobileNet. The actual inference time on a mobile device is also evaluated, with ShuffleNet achieving a ~13× speedup over AlexNet.ShuffleNet is an extremely efficient convolutional neural network (CNN) designed for mobile devices with limited computing power, such as 10-150 MFLOPs. The architecture introduces two novel operations: pointwise group convolution and channel shuffle, which significantly reduce computation cost while maintaining accuracy. Experiments on ImageNet classification and MS COCO object detection demonstrate that ShuffleNet outperforms other structures, achieving a lower top-1 error (7.8% absolute) than MobileNet on ImageNet at 40 MFLOPs. On an ARM-based mobile device, ShuffleNet achieves a ~13× actual speedup over AlexNet while maintaining comparable accuracy. The paper also explores the impact of group numbers and channel shuffle on performance, showing that larger group numbers and channel shuffle improve accuracy, especially for smaller networks. Additionally, ShuffleNet is evaluated on the MS COCO object detection task, showing superior performance compared to MobileNet. The actual inference time on a mobile device is also evaluated, with ShuffleNet achieving a ~13× speedup over AlexNet.