TL;DR
Most modern visual SLAM systems rely heavily on Bundle Adjustment (BA), which requires complex initialization heuristics and notoriously struggles with pure rotational or slow camera motion. This paper introduces "L-infinity SLAM," a novel pipeline that completely replaces BA by decoupling orientation estimation from position and map estimation. By using rotation averaging and globally optimal quasi-convex optimization, the proposed system drastically simplifies the SLAM pipeline and robustly handles challenging camera motions where traditional systems fail.
What Problem Does This Paper Address?
In feature-based monocular SLAM (like the famous ORB-SLAM), Bundle Adjustment (BA) is the mathematical engine. It simultaneously estimates the 6-degree-of-freedom (6DOF) camera trajectory and the 3D map (point cloud) by minimizing reprojection errors. However, BA is a highly non-linear, non-convex problem. This leads to two fundamental weaknesses:
- High System Complexity: Because BA only finds local minimums, it requires highly accurate initial guesses. SLAM systems must therefore constantly maintain and update a 3D map using a fragile web of heuristics (point culling, map aggregation, threshold tuning) just to keep BA from failing.
- The "SLAM Wiggle" Problem: Estimating 3D structure requires a baseline (translation between camera frames). If a camera moves very slowly or only rotates (e.g., panning a smartphone), BA cannot triangulate points. The math breaks down, causing the SLAM system to fail to initialize or lose tracking entirely.
The authors ask: Do we really need Bundle Adjustment? They propose an alternative optimization core that avoids these pitfalls entirely.
Methods
The authors propose L-infinity SLAM, which splits the SLAM problem into two distinct, mathematically friendlier steps:
- Rotation Averaging (Online Tracking): Instead of estimating pose and 3D points simultaneously, the real-time tracking thread only estimates camera orientations. It uses relative rotations (calculated from essential matrices) and averages them to find absolute camera orientations. This is fast and completely independent of 3D map points.
- The Known Rotation Problem (Background Mapping): Once camera orientations are known, finding the camera positions and 3D map points becomes the "Known Rotation Problem" (KRot). Instead of minimizing the sum of squared errors ($L_2$ norm) like BA, KRot minimizes the maximum reprojection error ($L_\infty$ norm).
- Why is this a big deal? The $L_\infty$ formulation is quasi-convex. This means it can be solved globally optimally. It does not need a careful initial guess, eliminating the need for complex map-maintenance heuristics.
Novel Loop Closure: Standard KRot is too slow for massive loop closures. The authors introduce a novel formulation that incorporates relative translation directions (extracted from the essential matrix) to constrain the problem, allowing it to solve loop closures efficiently over a sampled set of feature tracks.
Datasets
- Maptek Dataset: A real-world sequence (1833 frames) captured from a forward-looking camera mounted on a truck. It includes high-precision Inertial Navigation System (INS) data to provide absolute ground truth for camera poses.
- Smartphone Dataset: A custom video recorded by a pedestrian walking slowly and turning right. This was specifically captured to test the "low speed, high rotation" failure mode of standard SLAM.
Key Insights and Ideas
- Decoupling is a superpower: By separating rotation estimation from translation/mapping, the SLAM system doesn't need to constantly maintain a perfect 3D map just to figure out where the camera is looking. Mapping can be relegated to a lower-priority background thread.
- Immunity to pure rotation: Because the online tracking step relies on rotation averaging rather than 3D point triangulation, L-infinity SLAM can easily initialize and track a camera that is purely panning—a scenario that instantly breaks traditional BA-based SLAM.
- Global optimality removes heuristics: Because the KRot problem is quasi-convex, the solver will always find the best mathematical answer regardless of the starting point. This allows developers to strip away the messy, threshold-heavy heuristics usually required to "babysit" Bundle Adjustment.
Key Results
- Robustness over BA: On the Maptek dataset, a standard BA-SLAM implementation actually failed to complete the sequence (it disconnected due to outlier removal heuristics failing). L-infinity SLAM successfully completed the sequence, achieving highly accurate results (position error < 0.67m, rotation error < 0.83°).
- Surviving the Smartphone Test: On the pedestrian smartphone dataset, ORB-SLAM completely failed to initialize due to the low baseline and rotational motion. L-infinity SLAM successfully initialized and produced a clean 3D reconstruction.
- Speed: The incremental rotation averaging used in L-infinity SLAM proved to be an order of magnitude faster than incremental Bundle Adjustment. The novel loop closure formulation solved a 300-point loop closure in ~90 seconds.
Limitations and Future Work
- Sensitivity to Outliers: The authors acknowledge that minimizing the maximum error ($L_\infty$ norm) makes the math quasi-convex, but it also makes it inherently sensitive to outliers (if one point is wildly wrong, minimizing the maximum error will skew the whole map). While they note that robust filtering strategies exist to mitigate this, it remains a mathematical reality of the $L_\infty$ approach.
- Future Work: The authors hope this proof-of-concept will motivate the wider SLAM community to explore quasi-convex optimization as a viable, simpler alternative to the deeply entrenched Bundle Adjustment paradigm.
Summary updated