OpenVSLAM

Sumikura 2019 · Paper

One-line summary — A modular, open-source visual SLAM framework built as a callable library, with an ORB-based indirect pipeline and support for perspective, fisheye, and equirectangular (360-degree) camera models plus map save/load and localization.

Problem

Visual SLAM is essential for AR devices and autonomous robots, drones, and UAVs, but “conventional open-source visual SLAM frameworks are not appropriately designed as libraries called from third-party programs.” ORB-SLAM, LSD-SLAM, and DSO — the de facto standards — are research applications: hard to embed, hard to extend, limited to perspective cameras, and (per the paper’s comparison table) mostly unable to store/load maps. SfM packages like OpenMVG and OpenSfM already handled fisheye and equirectangular imagery; no real-time visual SLAM framework did.

Method & architecture

OpenVSLAM (MM ‘19) is a monocular/stereo/RGB-D system implemented in C++ (Eigen, OpenCV, g2o) whose contribution is architectural rather than algorithmic — well-known SLAM approaches encapsulated in separated components with clear APIs and extensive documented sample code.

Results

On a Core i7-7820HK laptop, compared against ORB-SLAM2:

Why it matters for SLAM

OpenVSLAM was one of the first SLAM codebases engineered for usability and integration rather than just benchmark accuracy, and it was widely adopted in research and industry as an embeddable component — especially for panoramic capture rigs no other open system supported. Its history also carries a lesson: concerns that code derived from GPL-licensed ORB-SLAM2 led to the original repository being taken down, and the community continued development as Stella-VSLAM. It remains a good entry point for studying a well-organised feature-based SLAM architecture.