History

Wang Xinyu 6c9bc7a4fb Update README.md		2020-04-09 21:49:22 +08:00
..
CMakeLists.txt	retinaface tested on 1080/trt7	2020-04-09 21:36:52 +08:00
common.h	add retinaface, unfinished	2020-04-01 21:19:57 +08:00
decode.cu	retinaface tested on 1080/trt7	2020-04-09 21:36:52 +08:00
decode.h	retinaface tested on 1080/trt7	2020-04-09 21:36:52 +08:00
plugin_factory.cpp	retinaface ssh, decode plugin	2020-04-08 14:42:02 +00:00
plugin_factory.h	retinaface ssh, decode plugin	2020-04-08 14:42:02 +00:00
README.md	Update README.md	2020-04-09 21:49:22 +08:00
retina_r50.cpp	retinaface tested on 1080/trt7	2020-04-09 21:36:52 +08:00

README.md

RetinaFace

Notice

Tested on TX2/TensorRT4 and GTX1080/TensorRT7
The pytorch implementation is biubug6/Pytorch_Retinaface, I forked it into wang-xinyu/Pytorch_Retinaface and add genwts.py

Run

1. generate retinaface.wts from pytorch implementation https://github.com/wang-xinyu/Pytorch_Retinaface

git clone https://github.com/wang-xinyu/Pytorch_Retinaface.git
// download its weights 'Resnet50_Final.pth', put it in Pytorch_Retinaface/weights
cd Pytorch_Retinaface
python detect.py --save_model
python genwts.py
// a file 'retinaface.wts' will be generated.

2. put retinaface.wts into tensorrtx/retinaface, build and run

git clone https://github.com/wang-xinyu/tensorrtx.git
cd tensorrtx/retinaface
// put retinaface.wts here
mkdir build
cd build
cmake ..
make
sudo ./retina_r50 -s  // build and serialize model to file i.e. 'retina_r50.engine'
wget https://github.com/TencentYoutuResearch/FaceDetection-DSFD/raw/master/data/worlds-largest-selfie.jpg
sudo ./retina_r50 -d  // deserialize model file and run inference.

3. check the images generated, as follows. result.jpg