1 Nanjing University 2 Robbyant, Ant Group 3 HKUST 4 Wuhan University 5 Zhejiang University
TL;DR Map2SLAM uses a frozen streaming 3D foundation model as a geometry engine, scheduling its causal and bidirectional inference to maintain a persistent, revisable SLAM state.
Map2SLAM keeps LingBot-Map frozen and shares Cached Frame Features across its operations. Initialization uses a Bidirectional Prefix to establish the Live Context; Tracking then follows incoming frames through Causal Continuation, extending the persistent trajectory.
The back end runs Bidirectional Prefix in a Fresh Context using the cached frame features. Loop-Pair Geometry verifies loop candidates and supplies constraints for online trajectory correction. Joint inference over overlapping segments supplies observations for Dense Optimization, which jointly refines keyframe poses and depths after the sequence.
A strong model matters.
So does how we use it.
Map2SLAM has one Front-end Tracking path and two Back-end optimization paths. The Front-end produces sequential observations online. Both Back-end paths use fresh Bidirectional Prefix inference on selected historical keyframes to obtain geometric observations. Online Trajectory Correction revises the trajectory during tracking, while Dense Geometric Refinement jointly refines keyframe poses and depths after the sequence.
Bidirectional Prefix · Causal Continuation · Keyframe Probe
Bounded tracking windows stabilize local motion estimates. Each probe yields pose, depth, and intrinsics; accepted probes commit their KV updates and provide sequential observations, while rejected probes roll back without replay.
Loop Retrieval · Pairwise Geometry · iPGO
A trained lightweight head maps cached Frame Encoder features to global keyframe descriptors. Descriptor similarity proposes loop candidates; fresh two-view Bidirectional Prefix verifies geometry, and iPGO corrects the trajectory during tracking.
Segment Scheduling · Segment Geometry · Dense Optimization
Sequential and loop segments select historical keyframes, fresh multi-view Bidirectional Prefix inference provides segment observations, and final pose–depth optimization jointly refines keyframe poses and depths.
Normalized timeline used only for trajectory inspection. While tracking continues, Map2SLAM (Ours) asynchronously corrects the trajectory using Incremental Pose Graph Optimization (iPGO).
Mean absolute trajectory error after global Sim(3) alignment for uncalibrated methods. Lower is better.
| Method | KITTI | VBR | Oxford Spires |
|---|---|---|---|
| LoGeR | 25.43 | 36.18 | 6.88 |
| LingBot-Map (S) | 26.30 | 26.71 | 8.14 |
| LingBot-Map (W) | 35.54 | 77.22 | 15.93 |
| HorizonStream | 17.91 | 18.18 | 15.66 |
| ABot-Recon | 13.56 | 9.97 | 5.36 |
| Ours | 7.56 | 5.15 | 3.07 |
Mean ATE(m) over 11 KITTI, 7 VBR, and 10 Oxford Spires sequences. Full per-sequence results and calibrated baselines are reported in the manuscript.
We train a lightweight head to convert cached Frame Encoder features into global descriptors, keeping the encoder frozen. The 1.8M-parameter head adds 0.94 ms per descriptor and reaches 88.2 mean R@1 across MSLS-val, Pitts250k-test, and SPED.
The system uses two NVIDIA RTX 5090D GPUs: one runs front-end tracking at over 28 FPS, while the other handles loop closure and bidirectional re-observation. Dense geometric refinement runs after the input sequence.
