New LeanGate network reduces computational costs in transformer-based SLAM by 85% through early frame filtering before expensive geometric processing

LeanGate is a lightweight feed-forward network that predicts geometric utility scores to assess whether frames contain novel geometry before heavy processing

The system bypasses over 90% of redundant frames, eliminating wasteful dense geometric decoding on frames unlikely to contribute to mapping

Evaluations show LeanGate reduces tracking FLOPs by more than 85% and achieves 5x end-to-end speedup on standard SLAM benchmarks

Works as a plug-and-play module compatible with Geometric Foundation Models (GFMs) used in modern monocular SLAM systems