面向设备端视觉识别的低功耗推理:一种量化友好型方案
计算机视觉与模式识别
2019-03-19 v1
摘要
IEEE低功耗图像识别挑战赛(LPIRC)是始于2015年的年度竞赛,旨在鼓励针对低延迟、低功耗计算机视觉系统的软硬件联合解决方案。2018年该竞赛的第一赛道聚焦于在固定推理引擎和硬件条件下的软件方案创新。这一决定使参赛者能够在线提交模型,而无需构建并携带定制硬件到现场,从而吸引了史上最大数量的提交。在多样化的方案中,获胜方案提出了一种面向MobileNets的量化友好型框架,在留出数据集上达到72.67%的准确率,在Google Pixel2手机单CPU核心上平均延迟为27ms,优于当时最佳的实时MobileNet模型。
引用
@article{arxiv.1903.06791,
title = {Low Power Inference for On-Device Visual Recognition with a Quantization-Friendly Solution},
author = {Chen Feng and Tao Sheng and Zhiyu Liang and Shaojie Zhuo and Xiaopeng Zhang and Liang Shen and Matthew Ardi and Alexander C. Berg and Yiran Chen and Bo Chen and Kent Gauen and Yung-Hsiang Lu},
journal= {arXiv preprint arXiv:1903.06791},
year = {2019}
}
备注
Accepted At The 2nd Workshop on Machine Learning on the Phone and other Consumer Devices (MLPCD 2)