RetinaFace Network MXNet Symbol=>ONNX=>TensorRT

1.Reference

• RetinaFace github: https://github.com/deepinsight/insightface/tree/master/RetinaFace
• RetinaFace arxiv: RetinaFace: Single-stage Dense Face Localisation in the Wild
• get pretrained model: BaiduDrive or Dropbox

2.Export ONNX Model

• export resnet50 model

python3 export_onnx.py

• export mobilenet 0.25 model

python3 export_onnx.py  --network mnet --imagenet ./model/mobilenet_0_25 --pretrained ./model/mnet.25 --transpose ./model/mnet_retinaface --onnx mnet_retinaface.onnx --input_shape 1 3 640 640

3.Build RetinaFace_trt Project

mkdir build && cd build
cmake ..
make -j

4.Run RetinaFace_trt

./RetinaFace_trt ../config.yaml ../samples

5.Inference Time Benchmark

resolution	MXNet Symbol	TensorRT
640x640		13.2ms

6.Results