mini-caffe

編譯必須vs2013 12

https://github.com/luoyetx/mini-caffe

調用demo：

#include <vector>
#include <algorithm>
#include <caffe/caffe.hpp>
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>

using namespace cv;
using namespace std;
using namespace caffe;

struct BBox {
  float x1, y1, x2, y2, score;
  int label;
};

int main(int argc, char* argv[]) {
  if (caffe::GPUAvailable()) {
    caffe::SetMode(caffe::GPU, 0);
  }

  Net net("../models/ssd/MobileNetSSD_deploy.prototxt");
  net.CopyTrainedLayersFrom("../models/ssd/MobileNetSSD_deploy.caffemodel");
  Mat img = imread("../r-fcn/004545.jpg");

  Profiler* profiler = Profiler::Get();
  profiler->TurnON();
  uint64_t tic = profiler->Now();

  const float kScoreThreshold = 0.5f;
  const char* kClassNames[] = { "__background__", "aeroplane", "bicycle", "bird", "boat",
                                "bottle", "bus", "car", "cat", "chair",
                                "cow", "diningtable", "dog", "horse",
                                "motorbike", "person", "pottedplant",
                                "sheep", "sofa", "train", "tvmonitor" };
  
  // preprocess
  int height = img.rows;
  int width = img.cols;
  Mat imgResized;
  cv::resize(img, imgResized, Size(300, 300));

  vector<Mat> bgr;
  cv::split(imgResized, bgr);
  bgr[0].convertTo(bgr[0], CV_32F, 1.f/127.5f, -1.f);
  bgr[1].convertTo(bgr[1], CV_32F, 1.f/127.5f, -1.f);
  bgr[2].convertTo(bgr[2], CV_32F, 1.f/127.5f, -1.f);

  // fill network input
  shared_ptr<Blob> data = net.blob_by_name("data");
  data->Reshape(1, 3, imgResized.rows, imgResized.cols);
  const int bias = data->offset(0, 1, 0, 0);
  const int bytes = bias * sizeof(float);
  memcpy(data->mutable_cpu_data() + 0 * bias, bgr[0].data, bytes);
  memcpy(data->mutable_cpu_data() + 1 * bias, bgr[1].data, bytes);
  memcpy(data->mutable_cpu_data() + 2 * bias, bgr[2].data, bytes);

  // forward
  net.Forward();

  // get output, shape is N x 7
  shared_ptr<Blob> result = net.blob_by_name("detection_out");
  const float* result_data = result->cpu_data();
  const int num_det = result->num();
  vector<BBox> detections;
  for (int k = 0; k < num_det; ++k) {
    if (result_data[0] != -1 && result_data[2] > kScoreThreshold) {
      // [image_id, label, score, xmin, ymin, xmax, ymax]
      BBox bbox;
      bbox.x1 = result_data[3] * width;
      bbox.y1 = result_data[4] * height;
      bbox.x2 = result_data[5] * width;
      bbox.y2 = result_data[6] * height;
      bbox.score = result_data[2];
      bbox.label = static_cast<int>(result_data[1]);
      detections.push_back(bbox);
    }
    result_data += 7;
  }

  // draw
  for (auto& bbox : detections) {
    cv::Rect rect(bbox.x1, bbox.y1, bbox.x2 - bbox.x1 + 1, bbox.y2 - bbox.y1 + 1);
    cv::rectangle(img, rect, cv::Scalar(0, 0, 255), 2);
    char buff[300];
    sprintf(buff, "%s: %.2f", kClassNames[bbox.label], bbox.score);
    cv::putText(img, buff, cv::Point(bbox.x1, bbox.y1), FONT_HERSHEY_PLAIN, 1, Scalar(0, 255, 0));
  }

  uint64_t toc = profiler->Now();
  profiler->TurnOFF();
  profiler->DumpProfile("./ssd-profile.json");

  LOG(INFO) << "Costs " << (toc - tic) / 1000.f << " ms";
  cv::imwrite("./ssd-result.jpg", img);
  cv::imshow("result", img);
  cv::waitKey(0);
  return 0;
}