Блог

Я пытаюсь использовать Deeplabv3.onnx модель в OpenCV DNN. Модель, которую я использую, была экспортирована из PyTorch. Несмотря на то, что я не получаю никаких ошибок компиляции или выполнения, реализация не дает ожидаемого сегментированного результата. Я думаю, что выходной большой двоичный объект из сети был неправильно декодирован, что привело к неправильным результатам сегментации. Я в основном использую OpenCV DNN Segmentation.cpp пример кода и немного изменен для предварительной обработки входного изображения перед его передачей в сеть. Было бы здорово, если бы вы могли посоветовать или исправить мой код. Заранее спасибо за ваше драгоценное время.

Segmention.cpp Код:

 #include <sstream>

#include <opencv2/dnn.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>

std::string keys =
    "{ help  h         |                                                   | Print help message. }"
    "{ model           | deeplabv3.onnx      | Path to input image or video file. Skip this argument to capture frames from a camera. }"
    "{ config          | <none>                                            | Path to model config file}"
    "{ input i         | opencv-samples/data/vtest.avi                                                  | Path to input image or video file. Skip this argument to capture frames from a camera. }"
    "{ device          |  0                                                | camera device number. }"
    "{ initial_width   | 256                                               | Preprocess input image by initial resizing to a specific width.}"
    "{ initial_height  | 256                                               | Preprocess input image by initial resizing to a specific height.}"
    "{ width           | 224                                               | Path to input image or video file. Skip this argument to capture frames from a camera. }"
    "{ height          | 224                                               |  }"
    "{ scale           | 1.0                                               | Scale of the input image }"
    "{ rgb             | true                                              | Path to input image or video file. Skip this argument to capture frames from a camera. }"
    "{ mean            | 0.485 0.456 0.406                                             | Path to input image or video file. Skip this argument to capture frames from a camera. }"
    "{ std             | 0.229 0.224 0.225                                       | Preprocess input image by dividing on a standard deviation.}"
    "{ framework f     |                                                   | Optional name of an origin framework of the model. Detect it automatically if it does not set. }"
    "{ classes         |                                                   | Optional path to a text file with names of classes. }"
    "{ colors          |                                                   | Optional path to a text file with colors for an every class. "
                                                                            "An every color is represented with three values from 0 to 255 in BGR channels order. }"
    "{ backend         | 0                                                 | Choose one of computation backends: "
                                                                            "0: automatically (by default), "
                                                                            "1: Halide language (http://halide-lang.org/), "
                                                                            "2: Intel's Deep Learning Inference Engine (https://software.intel.com/openvino-toolkit), "
                                                                            "3: OpenCV implementation }"
    "{ target          | 1                                                 | Choose one of target computation devices: "
                                                                            "0: CPU target (by default), "
                                                                            "1: OpenCL, "
                                                                            "2: OpenCL fp16 (half-float precision), "
                                                                            "3: VPU }";

using namespace cv;
using namespace dnn;

std::vector<std::string> classes;
std::vector<Vec3b> colors;

void showLegend();

void colorizeSegmentation(const Mat amp;score, Mat amp;segm);

int main(int argc, char** argv)
{
    CommandLineParser parser(argc, argv, keys);
    parser.about("Semantic segmentation deep learning networks using OpenCV.");

    int rszWidth = parser.get<int>("initial_width");
    int rszHeight = parser.get<int>("initial_height");
    float scale = parser.get<float>("scale");
    Scalar mean = parser.get<Scalar>("mean");
    Scalar std = parser.get<Scalar>("std");
    bool swapRB = parser.get<bool>("rgb");
    int inpWidth = parser.get<int>("width");
    int inpHeight = parser.get<int>("height");
    String model = parser.get<String>("model");
    String config = parser.get<String>("config");
    String framework = parser.get<String>("framework");
    int backendId = parser.get<int>("backend");
    int targetId = parser.get<int>("target");

#ifdef DNDEBUG
    if (argc == 1 || parser.has("help"))
    {
        parser.printMessage();
        return 0;
    }
#endif

    // Open file with classes names.
    if (parser.has("classes"))
    {
        std::string file = parser.get<String>("classes");
        std::ifstream ifs(file.c_str());
        if (!ifs.is_open())
            CV_Error(Error::StsError, "File "   file   " not found");
        std::string line;
        while (std::getline(ifs, line))
        {
            classes.push_back(line);
        }
    }

    // Open file with colors.
    if (parser.has("colors"))
    {
        std::string file = parser.get<String>("colors");
        std::ifstream ifs(file.c_str());
        if (!ifs.is_open())
            CV_Error(Error::StsError, "File "   file   " not found");
        std::string line;
        while (std::getline(ifs, line))
        {
            std::istringstream colorStr(line.c_str());

            Vec3b color;
            for (int i = 0; i < 3 amp;amp; !colorStr.eof();   i)
                colorStr >> color[i];
            colors.push_back(color);
        }
    }

    if (!parser.check())
    {
        parser.printErrors();
        return 1;
    }

    CV_Assert(!model.empty());
    //! [Read and initialize network]
    Net net = readNet(model,config);
    net.setPreferableBackend(backendId);
    net.setPreferableTarget(targetId);
    //! [Read and initialize network]

    // Create a window
    static const std::string kWinName = "Deep learning semantic segmentation in OpenCV";
    namedWindow(kWinName, WINDOW_NORMAL);

    //! [Open a video file or an image file or a camera stream]
    VideoCapture cap;
    if (parser.has("input"))
        cap.open(parser.get<String>("input"));
    else
        cap.open(parser.get<int>("device"));
    //! [Open a video file or an image file or a camera stream]

    // Process frames.
    Mat frame, blob;
    while (waitKey(1) < 0)
    {
        cap >> frame;
        if (frame.empty())
        {
            waitKey();
            break;
        }

        if (rszWidth != 0 amp;amp; rszHeight != 0)
        {
            resize(frame, frame, Size(rszWidth, rszHeight),0,0,INTER_NEAREST);
        }

        //! [Create a 4D blob from a frame]
        blobFromImage(frame, blob, scale, Size(inpWidth, inpHeight), mean, swapRB, false);
        //! [Create a 4D blob from a frame]

                // Check std values.
        if (std.val[0] != 0.0 amp;amp; std.val[1] != 0.0 amp;amp; std.val[2] != 0.0)
        {
            // Divide blob by std.
            divide(blob, std, blob);
        }

        //! [Set input blob]
        net.setInput(blob);
        //! [Set input blob]
        //! [Make forward pass]
        Mat score = net.forward();
        //! [Make forward pass]

        Mat segm;
        colorizeSegmentation(score, segm);

        resize(segm, segm, frame.size(), 0, 0, INTER_NEAREST);
        addWeighted(frame, 0.1, segm, 0.9, 0.0, frame);

        // // Put efficiency information.
        std::vector<double> layersTimes;
        double freq = getTickFrequency() / 1000;
        double t = net.getPerfProfile(layersTimes) / freq;
        std::string label = format("Inference time: %.2f ms", t);
        putText(frame, label, Point(0, 15), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 255, 0));

        imshow(kWinName, frame);
        if (!classes.empty())
            showLegend();
    }
    return 0;
}

void colorizeSegmentation(const Mat amp;score, Mat amp;segm)
{
    const int rows = score.size[2];
    const int cols = score.size[3];
    const int chns = score.size[1];

    if (colors.empty())
    {
        // Generate colors.
        colors.push_back(Vec3b());
        for (int i = 1; i < chns;   i)
        {
            Vec3b color;
            for (int j = 0; j < 3;   j)
                color[j] = (colors[i - 1][j]   rand() % 256) / 2;
            colors.push_back(color);
        }
    }
    else if (chns != (int)colors.size())
    {
        CV_Error(Error::StsError, format("Number of output classes does not match "
                                         "number of colors (%d != %zu)", chns, colors.size()));
    }

    Mat maxCl = Mat::zeros(rows, cols, CV_8UC1);
    Mat maxVal(rows, cols, CV_32FC1, score.data);
    for (int ch = 1; ch < chns; ch  )
    {
        for (int row = 0; row < rows; row  )
        {
            const float *ptrScore = score.ptr<float>(0, ch, row);
            uint8_t *ptrMaxCl = maxCl.ptr<uint8_t>(row);
            float *ptrMaxVal = maxVal.ptr<float>(row);
            for (int col = 0; col < cols; col  )
            {
                if (ptrScore[col] > ptrMaxVal[col])
                {
                    ptrMaxVal[col] = ptrScore[col];
                    ptrMaxCl[col] = (uchar)ch;
                }
            }
        }
    }

    segm.create(rows, cols, CV_8UC3);
    for (int row = 0; row < rows; row  )
    {
        const uchar *ptrMaxCl = maxCl.ptr<uchar>(row);
        Vec3b *ptrSegm = segm.ptr<Vec3b>(row);
        for (int col = 0; col < cols; col  )
        {
            ptrSegm[col] = colors[ptrMaxCl[col]];
        }
    }
}

void showLegend()
{
    static const int kBlockHeight = 30;
    static Mat legend;
    if (legend.empty())
    {
        const int numClasses = (int)classes.size();
        if ((int)colors.size() != numClasses)
        {
            CV_Error(Error::StsError, format("Number of output classes does not match "
                                             "number of labels (%zu != %zu)", colors.size(), classes.size()));
        }
        legend.create(kBlockHeight * numClasses, 200, CV_8UC3);
        for (int i = 0; i < numClasses; i  )
        {
            Mat block = legend.rowRange(i * kBlockHeight, (i   1) * kBlockHeight);
            block.setTo(colors[i]);
            putText(block, classes[i], Point(0, kBlockHeight / 2), FONT_HERSHEY_SIMPLEX, 0.5, Vec3b(255, 255, 255));
        }
        namedWindow("Legend", WINDOW_NORMAL);
        imshow("Legend", legend);
    }
}
 

Предварительно обученный код python для преобразования модели torchvision в модель Onnx:

 import os
import torch
import torch.onnx
from torch.autograd import Variable
from torchvision import models


def get_pytorch_onnx_model(original_model):
    # define the directory for further converted model save
    onnx_model_path = "models"
    # define the name of further converted model
    onnx_model_name = "deeplabv3_resnet101.onnx"

    # create directory for further converted model
    os.makedirs(onnx_model_path, exist_ok=True)

    # get full path to the converted model
    full_model_path = os.path.join(onnx_model_path, onnx_model_name)

    # generate model input
    generated_input = Variable(
        torch.randn(1, 3, 224, 224)
    )

    # model export into ONNX format
    torch.onnx.export(
        original_model,
        generated_input,
        full_model_path,
        verbose=True,
        input_names=["input"],
        output_names=["output"],
        opset_version=11
    )

    return full_model_path


def main():
    # initialize PyTorch ResNet-101 model
    original_model = models.segmentation.deeplabv3_resnet101(pretrained=True)

    # get the path to the converted into ONNX PyTorch model
    full_model_path = get_pytorch_onnx_model(original_model)
    print("PyTorch ResNet-100 model was successfully converted: ", full_model_path)


if __name__ == "__main__":
    main()