Scale Anchoring via Implicit Semantics and Sequence-Consistent Loop Closure for VGGT-Based SLAM
Organizations: Anyverse Dynamics
Abstract
Feed-forward geometric foundation models enable dense monocular reconstruction without requiring camera calibration, but independently reconstructed submaps can accumulate scale drift, and even a valid revisit can yield an inconsistent cross-submap measurement. We extend VGGT-SLAM 2.0 with fixed-root scale anchoring and short-sequence-assisted direct loop connections. SALAD selects an image pair between the initial submap and each eligible new submap; joint VGGT inference then uses learned implicit semantic-geometric cues to supply a scale-only anchor while preserving the original overlap chain. For loop closure, neighboring image pairs support a consistency check and translation-scale alignment. Each accepted candidate contributes one robust Sim(3) factor between existing center-frame nodes, without an auxiliary loop submap. A 17-sequence comparison lowers mean Sim(3) ATE on nine TUM sequences from 0.1018 to 0.0451 m relative to the VGGT-SLAM 2.0 baseline. Anchoring improves all seven no-loop module ablations. On three Clio scenes, a fixed-loop comparison favors the direct connection pipeline over the tested auxiliary-submap pipeline. A separate candidate audit exposes additional rejections of visually overlapping pairs by the short-sequence check. These comparisons distinguish the gains from long-range scale anchoring and loop connection choices from the limitations of candidate verification.
Figures & tables
| Dataset | Sequence | Baseline | Full |
|---|---|---|---|
| TUM | 360 | 0.068 | 0.093 |
| desk | 0.183 | 0.034 | |
| desk2 | 0.048 | 0.046 | |
| floor | 0.243 | 0.055 | |
| plant | 0.038 | 0.026 | |
| room | 0.209 | 0.052 |
| Sequence | B0 | B1 | B2 | Full | |
|---|---|---|---|---|---|
| TUM room | 0.088 | 0.085 | 0.056 | 0.052 | 3 |
| Corridor 1-1 | 12.848 | 6.468 | 12.848 | 6.468 | 0 |
| LaMAR HGE | 7.948 | 1.622 | 7.778 | 1.404 | 1 |
| Clio Apartment | 0.181 | 0.140 | 0.164 | 0.118 | 1 |
| Clio Cubicle | 0.073 | 0.043 | 0.042 | 0.033 | 2 |
| Clio Office | 2.120 | 1.500 | 2.156 | 1.508 | 1 |
| Reference strategy | Room | Corridor | ADVIO | LaMAR |
|---|---|---|---|---|
| No anchor | 0.088 | 12.848 | 9.882 | 7.948 |
| Fixed-root + middle frame | 0.090 | 7.239 | 6.942 | 2.902 |
| Any history + SALAD | 0.080 | 7.194 | 9.793 | 8.168 |
| Fixed-root + SALAD (ours) | 0.085 | 6.468 | 6.791 | 1.622 |
| Apartment | Cubicle | Office | |
| Candidate pool | 2 | 4 | 9 |
| Single-pair accepted | 2 | 2 | 2 |
| Short-sequence accepted | 1 | 2 | 1 |
| Anchor-only | 0.140 | 0.043 | 1.500 |
| Anchor + auxiliary-submap loop | 0.505 | 0.175 | 3.370 |
| Anchor + direct loop (Full) | 0.118 | 0.033 | 1.508 |
Simultaneous Localization and Mapping
Simultaneous localization and mapping estimates a robot’s motion while building a map of an unknown environment.
See all 210 papers