We did some modifications to ORB-SLAM2, to make it run efficiently with the data set:

  • we did NOT modify the „ORBextractor.minThFAST“
  • we modified the size of the images:
    • Since the images from the data set are quite big, we decided to resize them from 1032×772 to 640×480 and all the calibration parameters where modified according to this resize.
    • We therefore modified the source code: In ros_stereo.cc in GrabStereo before „if(do_rectify)“:
      cv::Mat img_left, img_right;
      cv::resize(cv_ptrLeft->image, img_left, cv::Size(640,480));
      cv::resize(cv_ptrRight->image, img_right, cv::Size(640,480));
  • we did modify the size of the ROS subscriber’s queue:
    • to 1000, to make sure we don’t lose any frame
    • We therefore modified the source code: In ros_stereo.cc:
    • message_filters::Subscriber<sensor_msgs::Image> left_sub(nh, "/hcru1/pt_stereo_rect/left/image", 1000);
      message_filters::Subscriber<sensor_msgs::Image> right_sub(nh, "/hcru1/pt_stereo_rect/right/image", 1000);


Therefore, the modified calibration for ORB-SLAM2 is:  SUPER_calibration_ORBSLAM2

NOTE: We did use the ROSBAGS for ORB-SLAM2. If you want us to upload a specific rosbag, please contact us. In the rosbags, SUPER has its old name HCRU.