Simple and Lightweight Human Pose Estimation

Zhe Zhang, Jie Tang and Gangshan Wu, Nanjing University, China

depthwise convolution and attention mechanism

Lightweight Pose Network (LPN)

LPN-50 can achieve 68.7 in AP score on the COCO test-dev set,

2.7M parameters and 1.0 GFLOPs, while the inference speed is 17 FPS on an Intel i7-8700K CPU

Simple baselines for human pose  estimation and tracking ECCV2018과 유사한 구조.

 

Attention Mechanism

Multi-context attention for human pose estimation, in Proceedings CVPR 2017, first HPE

Gcnet: Non-local networks meet squeeze-excitation networks and beyond, upgrade NLNet + GC + SENet

NLNet: non local network

GC : global context, retain long-range dependencies