Head and Shoulder Detection Base on MTCNN

use coco dataset to detect head and shoulder. This implements is base on MTCNN Pretrain model has been placed in models.

prepare

• download coco keypoints dataset
• python preprocess/coco.py --data-dir {your coco dataset } --anotation {anotation} -o {coco.feather} # collect keypoints and gen boundbox
• python preprocess/image_process.py -n {pnet,rnet,onet} --preprocess-path {./data/coco.feather} # gen data for one stage

train

• train pnet

python nets/net.py -n pnet -lr 0.002 -w 2

• train rnet

python nets/net.py -n rnet -lr 0.002 -w 2

• train onet

python nets/net.py -n onet -lr 0.002 -w 2

hard mining

python preprocess/hard_mining.py -n rnet python preprocess/hard_mining.py -n onet

test

python nets/test.py -p video

Warning

• Prediction is much more slower than expected in keras, but when predicts it on arm-rk3399, it only cost about 100ms totally.(python is really slow)
• how to improve the performace
  • batch norm
  • change the prediction of bound box, taking consideration of yolo v2/3.
  • cleaning data (it`s really important!!!!!. Our generator scripts exist a lot of noisy)
  • pruning model
  • attention (channel or feature attention)

current result

metric	pnet	rnet	onet
acc	94%	96.1%	98.5%