使用 MXNet C Predict API 实现基于 SSD 的目标检测

本文记录使用 MXNet 的 C Predict API 实现目标检测的过程。主要是参考 classification 的例子实现了 detection.

官网提供的 model 是 train phase 的 model, 因此，输入有 data 和 label 两个 Input Node, 在预测的时候，不需要 label Input Node, 在下载到官网的 model 之后需要使用官网的 deploy.py 把 model 转换一下，方法如下：

# 首先下载训练好的 model 文件, 例如 https://github.com/zhreshold/mxnet-ssd/releases/download/v0.6/resnet50_ssd_512_voc0712_trainval.zip 放到 model 中.
# in /path/to/mxnet/example/ssd
python deploy.py --network=resnet50 --data-shape 512

#include <stdio.h>

// Path for c_predict_api
#include <mxnet/c_predict_api.h>

#include <iostream>
#include <fstream>
#include <string>
#include <vector>

#include <opencv2/opencv.hpp>

const mx_float DEFAULT_MEAN = 117.0;

// Read file to buffer
class BufferFile {
 public :
    std::string file_path_;
    int length_;
    char* buffer_;

    explicit BufferFile(std::string file_path)
    :file_path_(file_path) {

        std::ifstream ifs(file_path.c_str(), std::ios::in | std::ios::binary);
        if (!ifs) {
            std::cerr << "Can't open the file. Please check " << file_path << ". \n";
            length_ = 0;
            buffer_ = NULL;
            return;
        }

        ifs.seekg(0, std::ios::end);
        length_ = ifs.tellg();
        ifs.seekg(0, std::ios::beg);
        std::cout << file_path.c_str() << " ... "<< length_ << " bytes\n";

        buffer_ = new char[sizeof(char) * length_];
        ifs.read(buffer_, length_);
        ifs.close();
    }

    int GetLength() {
        return length_;
    }
    char* GetBuffer() {
        return buffer_;
    }

    ~BufferFile() {
        if (buffer_) {
          delete[] buffer_;
          buffer_ = NULL;
        }
    }
};

void GetImageFile(const std::string image_file,
                  mx_float* image_data, const int channels,
                  const cv::Size resize_size, const mx_float* mean_data = nullptr) {
    // Read all kinds of file into a BGR color 3 channels image
    cv::Mat im_ori = cv::imread(image_file, cv::IMREAD_COLOR);

    if (im_ori.empty()) {
        std::cerr << "Can't open the image. Please check " << image_file << ". \n";
        assert(false);
    }

    cv::Mat im;

    resize(im_ori, im, resize_size);

    float mean_b, mean_g, mean_r;
    mean_b = 104.0;
    mean_g = 117.0;
    mean_r = 123.0;
    for(int i=0; i < im.cols; ++i){
      uchar* data = im.ptr<uchar>(i);
      for(int j=0; j< im.rows; ++j){
        image_data[i*im.cols+j+2] = static_cast<mx_float>(*data++) - mean_b;
        image_data[i*im.cols+j+1] = static_cast<mx_float>(*data++) - mean_g;
        image_data[i*im.cols+j] = static_cast<mx_float>(*data++) - mean_r;
      }
    }
}

int main(int argc, char* argv[]) {
    if (argc < 4) {
        std::cout << "Usage: ./detect symbol_path params_path image_path" << std::endl;
        return 0;
    }

    std::string test_file;
    test_file = std::string(argv[3]);

    // Models path for your model, you have to modify it
    std::string json_file = std::string(argv[1]);
    std::string param_file = std::string(argv[2]);

    BufferFile json_data(json_file);
    BufferFile param_data(param_file);

    // Parameters
    int dev_type = 1;  // 1: cpu, 2: gpu
    int dev_id = 0;  // arbitrary.
    mx_uint num_input_nodes = 1;  // 1 for feedforward
    const char* input_key[1] = {"data"};
    const char** input_keys = input_key;

    // Image size and channels
    int width = 512;
    int height = 512;
    int channels = 3;

    const mx_uint input_shape_indptr[2] = { 0, 4 };
    const mx_uint input_shape_data[4] = { 1, static_cast<mx_uint>(channels), static_cast<mx_uint>(height), static_cast<mx_uint>(width)};
    PredictorHandle pred_hnd = 0;

    if (json_data.GetLength() == 0 || param_data.GetLength() == 0) {
        return -1;
    }

    // Create Predictor
    MXPredCreate((const char*)json_data.GetBuffer(),
                 (const char*)param_data.GetBuffer(),
                 static_cast<size_t>(param_data.GetLength()),
                 dev_type, dev_id, num_input_nodes, input_keys, input_shape_indptr, input_shape_data, &pred_hnd);
    assert(pred_hnd);

    int image_size = width * height * channels;


    // // Read Image Data
    std::vector<mx_float> image_data = std::vector<mx_float>(image_size);
    GetImageFile(test_file, image_data.data(), channels, cv::Size(width, height));
    // // Set Input Image
    int64 start = cv::getTickCount();
    MXPredSetInput(pred_hnd, "data", image_data.data(), image_size);
    // // Do Predict Forward
    for(int i=0; i<20; ++i){
      MXPredForward(pred_hnd);
    }

    mx_uint output_index = 0;

    mx_uint *shape = 0;
    mx_uint shape_len;

    // Get Output Result
    MXPredGetOutputShape(pred_hnd, output_index, &shape, &shape_len);

    size_t size = 1;
    for (mx_uint i = 0; i < shape_len; ++i) size *= shape[i];

    std::vector<float> data(size);

    MXPredGetOutput(pred_hnd, 0, data.data(), size);
    int64 end = cv::getTickCount();
    double secs = (end-start)/cv::getTickFrequency();
    std::cout<<"time: " <<secs<<std::endl;
    assert(data.size() % 6 == 0);
    cv::Mat mat = cv::imread(test_file, 1);
    int orig_cols = mat.cols;
    int orig_rows = mat.rows;
    cv::resize(mat, mat, cv::Size(width, height));
    for(int i=0; i<data.size(); i+=6){
      if(data[i]<0) continue;
      int id = static_cast<int>(data[i]);
      float score = data[i+1];
      if(score < 0.5) continue;
      std::cout<<score<<std::endl;
      int xmin = static_cast<int>((data[i+2])*width);
      int ymin = static_cast<int>((data[i+3])*height);
      int xmax = static_cast<int>((data[i+4])*width);
      int ymax = static_cast<int>((data[i+5])*height);
      cv::rectangle(mat, cv::Point(xmin, ymin), cv::Point(xmax, ymax), cv::Scalar(255, 0, 0), 2);
    }
    cv::resize(mat, mat, cv::Size(orig_cols, orig_rows));
    cv::imshow(test_file, mat);
    cv::waitKey(0);
    return 0;
}

}