MTCNN head detection

Set up

Set up environment and copy C++ layer code to Caffe's source code tree.

$ export PYTHONPATH=/path/to/mtcnn-head-detection:$PYTHONPATH
$ export CAFFE_HOME=/path/to/caffe
$ pip install easydict
$ pip install lmdb
$ sh layers/copy.sh

Compile Caffe following its document.

Prepare data

Download dataset SCUT-HEAD. Unzip and put them in data directory.

Train

pnet

python jfda/prepare.py --net p --wider --worker 8
python jfda/train.py --net p --gpu 0 --size 128 --lr 0.05 --lrw 0.1 --lrp 5 --wd 0.0001 --epoch 25

rnet

Choose appropriate pnet caffemodel to generate prior for rnet, and edit cfg.PROPOSAL_NETS in config.py

python jfda/prepare.py --net r --gpu 0 --detect --wider --worker 4
python jfda/train.py --net r --gpu 0 --size 128 --lr 0.05 --lrw 0.1 --lrp 5 --wd 0.0001 --epoch 25

onet

Choose appropriate rnet caffemodel to generate prior for onet, and edit cfg.PROPOSAL_NETS in config.py

python jfda/prepare.py --net o --gpu 0 --detect --wider --worker 4
python jfda/train.py --net o --gpu $GPU --size 64 --lr 0.05 --lrw 0.1 --lrp 7 --wd 0.0001 --epoch 35

Test

python simpledemo.py

Note

1. Landmark alignment in original mtcnn is removed in this repo. Here only do object classification and bounding box regression.

2. Each convolutional layer kernel number in onet has reduced for faster network inference.

Result

pnet

rnet

onet

References

• A Convolutional Neural Network Cascade for Face Detection
• Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks
• MTCNN_face_detection_alignment
• py-faster-rcnn
• opencv-mtcnn