// the find
geohot/twitchslam
A toy implementation of monocular SLAM written while livestreaming
A monocular SLAM pipeline geohot built live on stream, covering feature extraction, pose estimation via essential matrices, keyframe-free tracking, and g2o bundle adjustment with a Pangolin 3D viewer. It's for people who want to see SLAM's moving parts laid bare in ~500 lines rather than wade into ORB-SLAM's C++.
The pipeline is small enough to read end to end in an afternoon and actually maps to the textbook SLAM stages (frame -> features -> pose -> map -> optimize). It ships real test videos plus ground-truth trajectories (freiburg, kitti) so you can sanity-check output against something concrete instead of eyeballing it. Uses proven libraries for the hard parts (g2opy for optimization, Pangolin for 3D) rather than reimplementing bundle adjustment badly.
Ships prebuilt .so binaries for g2opy and Pangolin pinned to Python 3.6 on Linux/macOS only — those wheels won't load on any current Python, and building g2opy/Pangolin from source yourself is its own multi-hour yak-shave. Last pushed 2022 with no CI and an author-written note that the accuracy test 'doesn't work reliably.' The README's own TODO list still has open items like frequent lockups and a REVERSE env var hack needed just to get initialization working, so this is stream-quality code, not something to build a product on.