Search code examples
c++opencvros

Why does cv::drawKeypoints resize my image?

I'm using OpenCV in C++ to process a cv::Mat before printing it to a ROS topic. For some reason cv::drawKeypoints messes up my result by virtually stretching it over the width beyond the image frame: example. The blob in the right topic represents the one on the top left in the left topic.

Here's my code:

image_transport::Publisher  pubthresh;
image_transport::Publisher  pubkps;
cv::SimpleBlobDetector detector;
void imageCallback(const sensor_msgs::ImageConstPtr& msg)
{
  cv::Mat mat = cv_bridge::toCvShare(msg, "bgr8")->image;
  cv::cvtColor(mat,mat, CV_BGR2GRAY );
  cv::threshold(mat,mat,35,255,0);
  std::vector<cv::KeyPoint> keypoints;
  detector.detect(mat, keypoints);
  cv::Mat kps;
  cv::drawKeypoints( mat, keypoints, kps, cv::Scalar(0,0,255), cv::DrawMatchesFlags::DRAW_RICH_KEYPOINTS );
  sensor_msgs::ImageConstPtr ithresh,ikps;
  ithresh = cv_bridge::CvImage(std_msgs::Header(), "mono8", mat).toImageMsg();
  ikps = cv_bridge::CvImage(std_msgs::Header(), "mono8", kps).toImageMsg();
  pubthresh.publish(ithresh);
  pubkps.publish(ikps);
}
int main(int argc, char **argv)
{
  ...
  image_transport::Subscriber sub = it.subscribe("/saliency_map", 1, imageCallback);
  ...
}

After the cv::drawKeypoints operation both cv::Mat are treated the same. According to the documentation the image shouldn't get resized either. What am I missing here?

Solution

  • Looks like your result image isn't grayscale but color image. Stretching here means, that each pixel becomes implicitly 3x the size in horizontal direction, because of having 3 channels, which are interpreted as grayscale values.

    So try to convert kps to grayscale before using your publishing stuff.

    cv::cvtColor(kps,kps, CV_BGR2GRAY );

    Or adjust the line 

    ikps = cv_bridge::CvImage(std_msgs::Header(), "mono8", kps).toImageMsg();

    to publish a bgr color image instead of "mono8". But I don't know how to use that code.