- 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.
- 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.
- 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.