I am most probably missing a point, either I am not using image_transport properly or there is a bug in the image_transport package.

The best alternative I have found now is the following one:

#include <ros/ros.h>
#include <opencv2/opencv.hpp>
#pragma GCC diagnostic ignored "-Wunused-parameter"
#include <cv_bridge/cv_bridge.h>
#pragma GCC diagnostic pop

int main(int argc,
         char** argv)
{
  ros::init(argc, argv, "image_publisher");
  ros::NodeHandle nh;

  std::string publish_topic("/webcam");
  ros::Publisher pub = nh.advertise<sensor_msgs::Image>(publish_topic, 5);
  ros::Publisher c_pub = nh.advertise<sensor_msgs::CompressedImage>(publish_topic + "/compressed", 5);

  cv::VideoCapture cap;
  if (!cap.open(0))
    return 1;

  ros::Rate loop_rate(25);
  while (nh.ok())
  {
    cv::Mat frame;
    cap >> frame;
    if (frame.empty())
      continue;

    sensor_msgs::ImagePtr msg = cv_bridge::CvImage(std_msgs::Header(), "bgr8", frame).toImageMsg();
    sensor_msgs::CompressedImagePtr c_msg =
        cv_bridge::CvImage(std_msgs::Header(), "rgb8", frame).toCompressedImageMsg();

    pub.publish(msg);
    c_pub.publish(c_msg);
    ros::spinOnce();
    loop_rate.sleep();
  }

  return 0;
}

I set two publishers: one for the raw image (bgr8 encoding) and one for the compressed image (rgb8 encoding) This way I get both the JPEG compressed images and the raw image with decent performance and the right colors.