ORB-SLAM3 is the third-generation visual SLAM system from the SLAMLab at the University of Zaragoza, released in 2020 under GPLv3. It was the first open-source library able to perform visual, visual-inertial, and multi-map SLAM with monocular, stereo, RGB-D, and fisheye-lens cameras in a unified framework. The system has two principal innovations: a feature-based tightly-coupled visual-inertial SLAM that uses Maximum-a-Posteriori estimation throughout, including IMU initialisation, reportedly making it 2-5x more accurate than previous approaches; and a multi-map (Atlas) system with a new place-recognition method that lets the system survive long periods of poor visual conditions by starting a fresh map and seamlessly merging it with prior sessions on revisit. On the EuRoC drone benchmark, stereo-inertial ORB-SLAM3 achieves average accuracy around 3.6 cm, and on the TUM-VI handheld dataset it reaches roughly 9 mm under fast motion — performance representative of AR/VR scenarios. The C++ codebase ships with ROS wrappers and is the standard reference baseline against which new SLAM methods are compared. It depends on OpenCV for image manipulation and runs in real time on commodity CPUs, making it suitable for mobile robots, drones, and AR devices.
Open-source feature-based SLAM library from University of Zaragoza. First system to perform visual, visual-inertial, and multi-map SLAM with monocular, stereo, RGB-D, and fisheye cameras. Reports 3.6 cm accuracy on EuRoC and 9 mm under quick hand-held motion on TUM-VI.
ORB-SLAM3 is the third-generation visual SLAM system from the SLAMLab at the University of Zaragoza, released in 2020 under GPLv3. It was the first open-source library able to perform visual, visual-inertial, and multi-map SLAM with monocular, stereo, RGB-D, and fisheye-lens cameras in a unified framework. The system has two principal innovations: a feature-based tightly-coupled visual-inertial SLAM that uses Maximum-a-Posteriori estimation throughout, including IMU initialisation, reportedly making it 2-5x more accurate than previous approaches; and a multi-map (Atlas) system with a new place-recognition method that lets the system survive long periods of poor visual conditions by starting a fresh map and seamlessly merging it with prior sessions on revisit. On the EuRoC drone benchmark, stereo-inertial ORB-SLAM3 achieves average accuracy around 3.6 cm, and on the TUM-VI handheld dataset it reaches roughly 9 mm under fast motion — performance representative of AR/VR scenarios. The C++ codebase ships with ROS wrappers and is the standard reference baseline against which new SLAM methods are compared. It depends on OpenCV for image manipulation and runs in real time on commodity CPUs, making it suitable for mobile robots, drones, and AR devices.