Face detection based on ChainerCV Faster-RCNN implementation

Modified ChainerCV example to train on WIDER FACE dataset for face detection instead of on PASCAL VOC.

The original ChainerCV example code is: https://github.com/chainer/chainercv/tree/master/examples/faster_rcnn

Usage

Tested on Windows 10 (64 bit), Python 3.5.3 (installed by Anaconda)

This code depends on:

Chainer (https://chainer.org/)
CuPy (https://cupy.chainer.org/)
ChainerCV (https://github.com/chainer/chainercv)
SciPy (https://www.scipy.org/; for importing MATLAB .mat file)

Library version:

Chainer 2.0.0
CuPy 1.0.0.1
ChainerCV 0.5.1

A pre-trained model is excluded from this repositiory. It will be downloaded automatically when you execute demo.py. Otherwise, download it manually from the next link and put it on trained_model/snapshot_model.npz.

http://nixeneko.2-d.jp/hatenablog/20170724_facedetection_model/snapshot_model.npz

Demo

If a path to pretorained model is not given, tained_model/snapshot_model.npz is used. Since the pretrained model has large file size (484 MB), it is excluded from this repository. It will be downloaded automatically when demo.py is executed.

python demo.py [--gpu <gpu>] [--pretrained_model <model_path>] <image>.jpg

Train code

First, download the WIDER FACE dataset from the website: http://mmlab.ie.cuhk.edu.hk/projects/WIDERFace/

Download these files and extract:

Wider Face Training Images
Wider Face Validation Images
Face annotations

Execute with specifing dataset directories and annotation files (.mat) for training and validation sets, such as:

python train.py --gpu=0 --train_data_dir="WIDER_train" --train_annotation="wider_face_split/wider_face_train.mat" --val_data_dir="WIDER_val" --val_annotation="wider_face_split/wider_face_val.mat"

Or edit codes specifying dataset paths in train.py and execute:

python train.py --gpu 0

test_dataset.py

Show ground truth and predicted result on the images from validation dataset. Requires OpenCV Python bindings to execute.

Edit test_detection.py about dataset paths and run:

python test_detection.py --gpu 0

Green rectangles: ground truth.
Blue rectangles: predicted result (the less score has the lighter color)

Esc key to exit. Any other keys to the next image.

eval.py

Returns average precision. (0.362059 for trained_model/snapshot_model.npz)