Monocular SLAM often stumbles when objects in a scene move independently of the camera. Many recent systems try to identify such dynamic regions by estimating per-pixel uncertainty from rendering residuals or multiview feature inconsistency. The authors of FLoTS-SLAM argue that those cues lack direct motion evidence and can degrade when reconstruction is inaccurate or features are ambiguous.

Their proposed system instead derives uncertainty from the residual between optical flow and the flow induced by camera motion, which directly reflects object motion. Because the estimate adapts to frame alignment accuracy, it can reduce false detections and produce accurate pixel-level masks of dynamic regions. The system then uses high-confidence pixels to estimate camera poses, and those poses guide a joint refinement of poses and depth while reprojecting dynamic pixels, limiting the influence of object motion on both.

On benchmarks, FLoTS-SLAM improved mean dynamic mask IoU by 35% over a state-of-the-art baseline on DAVIS, and reduced average absolute trajectory error by 59% on Bonn and 19% on TUM relative to its SLAM backbone. The paper also provides a quantitative evaluation of dynamic masks and analyzes how those masks affect downstream tracking and reconstruction of dynamic objects. Source code is planned to be released.