ORB-SLAM3

Details of changes between the different versions.

OpenCV static matrices changed to Eigen matrices. The average code speed-up is 16% in tracking and 19% in mapping, w.r.t. times reported in the ORB-SLAM3 paper.
New calibration file format, see file Calibration_Tutorial. Added options for stereo rectification and image resizing.
Added load/save map functionalities.
Added examples of live SLAM using Intel Realsense cameras.
Fixed several bugs.

RGB-D compatibility: the RGB-D examples have been adapted to the new version.
Kitti and TUM dataset compatibility: these examples have been adapted to the new version.
ROS compatibility: updated the old references in the code to work with this version.
Config file parser: the YAML file contains the session configuration, a wrong parametrization may break the execution without any information to solve it. This version parses the file to read all the fields and give a proper answer if one of the fields have been wrongly deffined or does not exist.
Fixed minor bugs.

Initial release. It has these capabilities:

Multiple-Map capabilities: it is able to handle multiple maps in the same session and merge them when a common area is detected with a seamless fussion.
Inertial sensor: the IMU initialization takes 2 seconds to achieve a scale error less than 5% and it is reffined in the next 10 seconds until it is around 1%. Inertial measures are integrated at frame rate to estimate the scale, gravity and velocity in order to improve the visual features detection and make the system robust to temporal occlusions.
Fisheye cameras: cameras with wide-angle and fisheye lenses are now fully supported in monocular and stereo.