ycdhqzhiai/gazetensorrt

gaze estimation tensorrt

GazeTensorRT

gaze estimation tensorrt

run

• 生成onnx模型和wts文件

git clone https://github.com/ycdhqzhiai/Gaze-PFLD
cd Gaze-PFLD
python genwts.py
#同时生成onnx方便查看网络结构
python export_onnx.py

• 编译运行

mkdir build
cd build
cmake ..
make -j4
#1.得到engin
./gaze-pfld -s
#2 推理运行
./gaze-pfld -d

• 耗时

框架(GTX3090)	inference(ms)
tensorrt	0.47
torch	13.7
onnxruntime	3.8~270(不知道为什么)

• 关于误差
  对pth onnx engine三个模型都进行了测试，gaze部分，pth和engine基本一致，误差在小数点后第四位，但是landmarks部分误差较大，在小数点后第三位，onnx与engine也有一定误差，基本也在小数点后第三位，并且很奇怪的是，pth模型和onnx模型推理的结果也有一定误差，在小数点后第三位，令人不解

Reference resources

• 1.https://github.com/wang-xinyu/tensorrtx
  
• 2.https://github.com/ycdhqzhiai/Gaze-PFLD

gazetensorrt's Issues

Data pre-processing

I am using a real mobile device to capture the data.

Using opencv for greyscale processing and normalisation, the image is very different from pre_img.png.

My grayed-out image, which is not as detailed as the one you provided. For example, pupil detail is completely lost.

I would like to ask if there is anything special about the image pre-processing of this dataset.Or is there any way to keep my captured images consistent with the training set.

BR.