pywin / Optim_Hopenet

Deep Head(face) Pose Estimation

Optim_Hopenet Head Pose Estimation

Environment

• mxnet 1.5.0
• gluon-cv 0.5.0

Optim_net

Backbone: MobileNetv2, MobileNetv3 (Implement in gluon-cv)

Preparation

assume you are in the directory $Optim_Hopenet/.

1. download 300W_LP and AFLW_2000-3D
2. unzip
3. get 'pitch, yaw ,roll ' from annotation file

# change the dataset path to your path
python data/gen_pose.py

Train

python train.py --bs 128 --lr 0.001 --alpha 1 --lr_type cos --version small --width_mult 1 --use_fc 1 --net v3 --gpu 0 --prefix test

Backone	MAE(alpha=1)	MAE(alpha=2)	Mb
MobileFaceNet	6.760	6.876	4.1
MobileNetv2	6.510	6.549	9.8
MobileNetv3 small	6.660	6.706	7.5
MobileNetv3 large	6.293	6.145	17.5

Test

python test.py --test_type image --image test_res/test.jpg

Convert2Caffe

test: python mxnet2caffe/inference.py

References:

@InProceedings{Ruiz_2018_CVPR_Workshops,
author = {Ruiz, Nataniel and Chong, Eunji and Rehg, James M.},
title = {Fine-Grained Head Pose Estimation Without Keypoints},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2018}
}