// the find
gaoxiang12/ORB-YGZ-SLAM
A fork of ORB-SLAM2 that swaps in SVO-style direct tracking for the feature-matching step to get roughly 3x the speed, and adds monocular visual-inertial SLAM (VIO) on top. Aimed at robotics/CV researchers who already know the ORB-SLAM2 codebase and want a faster variant with IMU fusion, not a drop-in production SLAM library.
Hybrid tracking (ORB features + SVO-style direct alignment) is a real engineering idea, not just a rebrand, and the claimed 3x speedup with comparable accuracy is plausible given how expensive ORB feature matching is. Full IMU preintegration is implemented from scratch in src/IMU (NavState, IMUPreintegrator, g2o types for the VIO factors) rather than bolted on. It inherits ORB-SLAM2's mature bones: DBoW2 loop closure, g2o optimization, Sim3Solver for monocular scale, so the core SLAM pipeline isn't reinvented. Ships EuRoC/KITTI/TUM example runners plus Python ATE evaluation scripts, so you can actually reproduce a number instead of taking the README's word for it.
Dead since July 2017: no commits since, and the dependency list (Qt4, libcholmod3.0.6, old Pangolin/g2o APIs) means getting this to compile on anything current is its own multi-hour project before you evaluate the SLAM itself. The README is upfront that the direct-tracking robustness is weak and that it fails on harder EuRoC sequences (MH04/MH05, V102/V103, V2xx aren't claimed to pass) — so the speedup comes with a real accuracy/robustness tradeoff, not a free lunch. Licensing is murky: thirdparty folders carry their own GPL/BSD licenses but it's unclear what covers the YGZ-specific VIO/direct-tracking code, and the README explicitly says to email the author for commercial use. No tests, no CI, and documentation beyond the README is a single informal Note.md dev log — this is research code you debug by reading source, not something you integrate against a stable API.