291

这一模型和OpenPose高度相关，同时特征模型能与多种框架相关联。论文的作者提供了一种自下而上的方法，对多人的姿态进行实时估计，不需要用任何人物探测器。这种方法运用了一种非参数表示，我们称为Part Affinity Fields（PAFs），用它可以学习将图中人物和其身体部位联系到一起。有关该技术的具体细节和理论，可以阅读原文。

环境较难搭建
- CUDA 7.X
- Caffe 2016
- Matlab

Reference

[ ] Github repo
[ ] 论文

Training with mscoco2014

sudo apt install octave liboctave-dev texlive-lang-polish

Get Dataset

source getData.sh

convert the annotation format from json to mat

obatin the mask images for unlabeled person

generate a json file that contain raw informations needed for training.

generate LMDB.

python genLMDB.py

LMDB - Lightning Memory-Mapped Database

You can also download LMDB for the COCO dataset (189GB file) by: source get_lmdb.sh

Build pycaffe with caffe_train

the caffe is too old (2016).

make pycaffe

junxnone / tio

Realtime Multi-Person Pose Estimation #296

291

Reference

Training with mscoco2014

Get Dataset

convert the annotation format from json to mat

obatin the mask images for unlabeled person

generate a json file that contain raw informations needed for training.

generate LMDB.

Build pycaffe with caffe_train

295