Lite-HRNet: A Lightweight High-Resolution Network

Introduction

@inproceedings{Yulitehrnet21,
  title={Lite-HRNet: A Lightweight High-Resolution Network},
  author={Yu, Changqian and Xiao, Bin and Gao, Changxin and Yuan, Lu and Zhang, Lei and Sang, Nong and Wang, Jingdong},
  booktitle={CVPR},
  year={2021}
}

Results and models

Results on COCO val2017 with detector having human AP of 56.4 on COCO val2017 dataset

Arch Input Size #Params FLOPs AP AP50 AP75 AR AR50
Lite-HRNet-18 256x192 1.1M 205.2M 0.648 0.867 0.730 0.712 0.911
Lite-HRNet-18 384x288 1.1M 461.6M 0.676 0.878 0.750 0.737 0.921
Lite-HRNet-30 256x192 1.8M 319.2M 0.672 0.880 0.750 0.733 0.922
Lite-HRNet-30 384x288 1.8M 717.8M 0.704 0.887 0.777 0.762 0.928

Results on MPII val set.

Arch Input Size #Params FLOPs Mean Mean@0.1
Lite-HRNet-18 256x256 1.1M 273.4M 0.854 0.295
Lite-HRNet-30 256x256 1.8M 425.3M 0.870 0.313