SURE-Map: Self-Correcting Streaming Geometric Foundation Model
Mingkai Liu, Hao Zhao, Xingxing Zuo
cs.CV
2026-09-15
On a frozen LingBot-Map backbone, SURE-Map adds cross-view uncertainty and multi-timescale scale recalibration, cutting KITTI ATE from 24.00 m to 17.24 m, or 15.17 m with loops.
Streaming geometric foundation models are starting to replace classical SLAM: each arriving frame yields pose and dense depth without explicit matching or a heavy back-end. The cost is bounded context. Dynamic objects, weak texture, and repetition produce local pose and geometry errors, and those errors accumulate into scale drift on long trajectories.
The usual alternatives are uncomfortable. SLAM hybrids fold feed-forward priors into keyframes and pose graphs; consistency improves, latency does not. Offline feed-forward models see a full sequence or large chunks and cannot run strictly online. Compact streaming systems such as LingBot-Map and HorizonStream keep local scale reasonably stable, but they lack cross-segment constraints, so scale still walks between segments.
SURE-Map sits on a frozen LingBot-Map backbone and trains only a cross-view geometric uncertainty head. From the backbone's adjacent-frame pose and depth one can compute pose-depth-induced optical flow in closed form. The head reads geometric tokens from two frames and emits per-pixel log-variances on the horizontal and vertical axes. Training uses TartanAir ground-truth flow residuals under a negative log-likelihood loss, so large residuals receive larger uncertainty. At inference the head never predicts a flow field; it only scores whether correspondences are trustworthy.
That map is used twice. First, point filtering: only pixels with valid depth and low scalar uncertainty are back-projected, which drops floaters in dynamic or ambiguous regions. The backbone's native depth confidence mainly reflects single-view geometry; in the filtering ablation it deletes valid structure. Second, local translation optimization: rotations stay locked, point-to-plane residuals are weighted by propagated uncertainty, and a few Gauss-Newton steps update translation. Locking rotation keeps scale and correspondence residuals from multiplying into rotational error.
Local edits cannot stop slowly accumulating scale. A slower path maintains a sparse keyframe window by stride and induced-flow magnitude, runs one full-attention inference over the window, aligns window depths to streaming depths by inverse-depth fitting, and compares relative translation lengths to obtain a segment scale. Recent edges keep their rotations and update translation only, fusing the scale-recalibrated translation with the local estimate through a hyperparameter μ. Optional loop closure retrieves revisits with cached DINOv2 features and feeds them to a pose graph.
Trajectory accuracy is measured on KITTI, Oxford Spires, and VBR; dense reconstruction on Neural RGB-D and 7-Scenes. KITTI mean ATE-RMSE: LingBot-Map 24.00 m, HorizonStream 22.38 m, SURE-Map 17.24 m, 15.17 m with loop closure. Oxford goes from 5.11 to 4.74, 4.63 with loops. VBR goes from 31.37 to 28.58, 22.12 with loops. Ablation: dropping scale recalibration restores 5.11 / 24.00 / 31.37, essentially the frozen backbone.
On dense reconstruction, Neural RGB-D F1 moves from LingBot-Map's 65.10 to 66.20 and accuracy from 0.074 m to 0.067 m; completeness slips from 0.030 m to 0.035 m. Filtering with depth confidence worsens Chamfer from 0.051 to 0.084. On KITTI sequence 04, throughput falls from 11.68 to 9.17 FPS, a 15 to 30 ms add-on per frame.
| Method | KITTI ATE | Oxford ATE | VBR ATE |
| LingBot-Map | 24.00 | 5.11 | 31.37 |
| HorizonStream | 22.38 | 8.87 | 29.68 |
| SURE-Map | 17.24 | 4.74 | 28.58 |
| SURE-Map + LC | 15.17 | 4.63 | 22.12 |
A streaming geometry model on long horizons has to notice when its own step is untrustworthy. Cross-view uncertainty asks whether pose plus depth induce consistent pixel correspondences, which is closer to association quality than single-view depth confidence. Multi-timescale correction keeps the fast path cheap and borrows long-range evidence only periodically. For teams already on a LingBot-Map-style backbone, this is a plug-in correction layer rather than a larger streaming model to retrain.
The 7 m KITTI drop is mostly scale recalibration. Oxford only moves 0.37 m. Gains are uneven.
The backbone is frozen, so correction cannot exceed LingBot-Map's local predictions. The uncertainty head is trained only on TartanAir ground-truth flow; calibration may drift on driving and indoor data. Local optimization locks rotation, so rotational drift waits for loop closure. On VBR, 28.58 m without loops is still worse than HorizonStream with loops at 22.87 m; loop quality can reorder the long-horizon ranking. Filtering costs some completeness. There is no standalone limitations section, and no quantitative detection rate for dynamic objects.