the retinaface is not right

[xinsuinizhuan]

I run the retinafece, it show this, so it is not right

[ttanzhiqiang]

this is my result

you use which model,and you show you param in you project.

[xinsuinizhuan]

this is my result

you use which model,and you show you param in you project.

int main_RetinaFaceDectector() { RetinaFaceDectector m_RetinaFaceDectector; Config m_config; m_config.onnxModelpath = "E:\comm_Item\Item_done\onnx_tensorrt_pro\onnx_tensorrt_project-main\model\mxnet_onnx_tensorrt_retinaface\mnet.25-512x512-batchsize_1.onnx"; m_config.engineFile = "E:\comm_Item\Item_done\onnx_tensorrt_pro\onnx_tensorrt_project-main\model\mxnet_onnx_tensorrt_retinaface\mnet.25-512x512-fp32_batchsize_1.engine"; m_config.calibration_image_list_file = "E:\comm_Item\Item_done\onnx_tensorrt_pro\onnx_tensorrt_project-main\model\mxnet_onnx_tensorrt_retinaface\image\"; m_config.maxBatchSize = 1; m_config.mode = 2; m_config.calibration_width = 512; m_config.calibration_height = 512; m_config.conf_thresh = 0.2; m_config.m_NMSThresh = 0.2; m_RetinaFaceDectector.init(m_config); std::vector batch_res; std::vector batch_img; std::string filename = "E:\comm_Item\Item_done\onnx_tensorrt_pro\onnx_tensorrt_project-main\model\mxnet_onnx_tensorrt_retinaface\image\img.jpg"; cv::Mat image = cv::imread(filename); //test1 batch_img.push_back(image);

float all_time = 0.0;
time_t start = time(0);
Timer timer;
int m = 1000;
for (int i = 0; i < m; i++)
{
    //timer.reset();
    clock_t start, end;
    timer.reset();
    m_RetinaFaceDectector.detect(batch_img, batch_res);
    //double t = timer.elapsed();
    double t = timer.elapsed();
    std::cout << i << ":" << t << "ms" << std::endl;
    if (i > 0)
    {
        all_time += t;
    }
}
std::cout << m << "次 time:" << all_time << " ms" << std::endl;
std::cout << "1次 time:" << all_time / m << " ms" << std::endl;
std::cout << "FPS::" << 1000 / (all_time / m) << std::endl;
//disp
for (int i = 0; i < batch_img.size(); ++i)
{
    for (const auto& r : batch_res[i])
    {
        std::cout << "batch " << i << " prob:" << r.prob << " rect:" << r.rect << std::endl;
        cv::rectangle(batch_img[i], r.rect, cv::Scalar(255, 0, 0), 2);
        std::stringstream stream;
        stream << std::fixed << std::setprecision(2)  << "  score:" << r.prob;
        cv::putText(batch_img[i], stream.str(), cv::Point(r.rect.x, r.rect.y - 5), 0, 0.5, cv::Scalar(0, 0, 255), 2);
        for (size_t j = 0; j < 5; j++) {
            cv::Point2f pt = cv::Point2f(r.m_FacePts.x[j], r.m_FacePts.y[j]);
            cv::circle(batch_img[i], pt, 1, cv::Scalar(0, 255, 0), 2);
        }
    }
    cv::namedWindow("image" + std::to_string(i), cv::WINDOW_NORMAL);
    cv::imshow("image" + std::to_string(i), batch_img[i]);
    //cv::imwrite("E:\\comm_Item\\Item_done\\onnx_tensorrt_pro\\onnx_tensorrt_project-main\\model\\mxnet_onnx_tensorrt_retinaface\\result\\image"  + std::to_string(i) +".png",batch_img[i]);
}
cv::waitKey();
return 0;

}

[xinsuinizhuan]

I use the defalut param model, the mnet.25-512x512-batchsize_1.onnx, this model should not right, i will try the retinaface_r50_v1-512x512-batchsize_1 model.

[ttanzhiqiang]

m_config.mode = 2 this mode is int8.you should changed m_config.mode = 0.int8 model should calibration.look https://github.com/ttanzhiqiang/onnx_tensorrt_project/blob/main/README.md .support FP32(m_config.mode = 0),FP16(m_config.mode = 1),INT8(m_config.mode = 2)

[xinsuinizhuan]

m_config.mode = 2 this mode is int8.you should changed m_config.mode = 0.int8 model should calibration.look https://github.com/ttanzhiqiang/onnx_tensorrt_project/blob/main/README.md .support FP32(m_config.mode = 0),FP16(m_config.mode = 1),INT8(m_config.mode = 2)

this model is int8? support FP32(m_config.mode = 0),FP16(m_config.mode = 1),INT8(m_config.mode = 2) then m_config.mode = 2 is right?

[ttanzhiqiang]

int8 need calibration image data.you should look int8 theory.like https://zhuanlan.zhihu.com/p/58182172

[xinsuinizhuan]

I see. tiny_tensorrt_onnx.vcxproj items all works. But tiny_tensorrt_dyn_onnx.vcxproj the model of model\yolov5\yolov5-v5_weight missing, i don't find them. Could share with me?

[ttanzhiqiang]

you can use yolov5l.sim.onnx.