MAC-VO

Why we included this project

MAC-VO is a learning-based stereo visual odometry system that estimates camera motion from a stereo image feed. It pairs a learned frontend that produces per-frame covariance with a pose-graph backend, and that uncertainty-aware design earned it the ICRA 2025 Best Paper award. The repo includes pretrained weights, Docker and ROS-2 integrations, per-sequence configs, plus dense mapping and a mixed-precision fast mode that runs around 12 fps at 480x640. There are evaluation and plotting scripts for trajectory error, so you can reproduce results on the provided demo sequence before adapting it to your own rig. It's research code for engineers building or tuning SLAM pipelines, not a turnkey product, but the extension documentation makes it a workable starting point for your own learning-based odometry work.