MTCNN NCNN Implementation

the application can be run on X86 and ARM platform with Neon optimization.

How to build

clone the repo

git clone --recursive https://github.com/infinivision/mtcnn_ncnn.git

build ncnn

ncnn as submodule for the main repo
to support mtcnn. I update the ncnn code. the repo addr
for the build and install please refer to the ncnn wiki page

copy the ncnn lib and include headers to the main repo folder

# by default the ncnn was installed in ncnn/build/install folder
mkdir -p lib/ncnn
mkdir -p include/ncnn
cp 3rdparty/ncnn/build/install/lib/libncnn.a lib/ncnn/
cp 3rdparty/ncnn/build/install/include/* include/ncnn/

build the mtcnn main repo

mkdir -p build
cd build
cmake ..
make

how to run

test picture

cd bin
./test_picture ../models/ncnn/ ../images/1.jpg

test video

cd bin
./test_video ../models/ncnn 0

benchmark

Test data

we use the WIDER FACE validate dataset as the test data

run benchmark

cd bin
./benchmark ../models/ncnn ../images

Test Result

Platform	CPU Cores	Memory	total images	detected	min time(ms)	max time(ms)	avg time(ms)
MacOS Sierra	8	16GB	3226	2073	8.068	287.446	47.03
firefly 3399	6	2GB	3226	2084	25.541	4437.42	351.578
firefly 3399(Neon)	6	2GB	3226	2084	10.578	1325.09	157.908