Stabilized Prompt-to-Geometry Interface for Robust Dynamic Visual–Inertial Odometry.
1Electronic Information School, Wuhan University 2College of Geodesy and Geomatics, Shandong University of Science and Technology †Corresponding author
Dynamic visual–inertial odometry (VIO) requires reliable suppression of motion-corrupted measurements, yet prior semantic-assisted approaches depend on category-limited segmenters and degrade under partial occlusion. Promptable foundation segmentation models offer category-agnostic dynamic parsing, but their effectiveness in VIO depends critically on the temporal stability of input prompts — a factor largely overlooked in existing pipelines. When prompts derived from raw detection are jittery or intermittent under occlusion, the resulting masks flicker across frames, destabilizing geometric estimation.
We propose STAG-VIO, which formulates dynamic robustness as a perception-to-geometry interface stabilization problem. We introduce uncertainty-adaptive multi-object tracking that models prompt generation as state estimation with bounded noise adaptation, producing temporally coherent box prompts. These stabilized prompts drive a lightweight foundation segmenter whose masks undergo geometry-oriented morphological refinement to establish conservative safety margins. A constraint-budget-aware feature redistribution strategy preserves well-conditioned static measurements when dynamic regions dominate the view.
Experiments on VIODE and OpenLORIS-Scene show consistent gains over state-of-the-art baselines. Ablation confirms that prompt stabilization is the single most impactful component, reducing trajectory error by up to 83%.
Noise introduced at any stage propagates and amplifies downstream: jittery detections → flickering prompts → inconsistent masks → corrupted geometric constraints. STAG-VIO inserts an explicit stabilization mechanism at every transition of the chain Dt → B̂t → Mreft → Pstatict → xt.
Prompt generation is modeled as state estimation with bounded noise adaptation. A constant-velocity Kalman filter with an erf-based, asymmetric trust mechanism adapts the measurement noise R from innovation statistics: trust consistent detections, downweight deviant ones — bridging detection gaps under occlusion.
Stable prompts drive a lightweight promptable segmenter (MobileSAM). A deliberately conservative “heavy-dilate, light-erode” refinement over-envelopes moving objects, keeping motion-blurred edge features out of the estimator — while temporally coherent prompts enable safe encoder embedding reuse for a 23% speedup.
Mask-gated feature filtering removes corrupted residuals, then ANMS-based redistribution replenishes spatially well-distributed static features up to a target budget — keeping visual–inertial optimization well-conditioned even when dynamic regions dominate the view.
Across the full none→high dynamic sweep, STAG-VIO stays within 0.094–0.246 m ATE in every environment — a spread under 2.2× — while every baseline exhibits at least one catastrophic failure or divergence.
| Method | City day | City night | Parking lot | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| none | low | mid | high | none | low | mid | high | none | low | mid | high | |
| ORB-SLAM3 | 2.179 | 5.301 | 2.204 | * | 0.176 | * | * | * | 0.287 | 2.897 | 6.742 | 7.482 |
| VINS-Fusion | 0.203 | 0.148 | 0.261 | 0.321 | 0.319 | 0.368 | 0.433 | 0.490 | 0.120 | 0.113 | 0.161 | 1.201 |
| VINS-Mono | 0.186 | 0.237 | 0.263 | 3.169 | 0.314 | 0.436 | 0.727 | 0.575 | 0.102 | 0.109 | 2.915 | 4.933 |
| VINS-Mah | 0.149 | 0.193 | 0.222 | 0.264 | 0.417 | 0.610 | 0.651 | 0.556 | 0.136 | 0.151 | 0.246 | 0.128 |
| RP-VIO | 0.382 | 0.229 | 0.435 | 0.536 | 0.263 | 0.509 | 0.652 | 0.577 | 0.981 | 1.334 | 0.375 | 0.713 |
| RD-VIO | 0.305 | 0.148 | 0.187 | 1.221 | 0.305 | 0.871 | 0.635 | 0.560 | 0.299 | 0.359 | 0.312 | 0.527 |
| Dyna-VINS | 0.349 | 0.330 | 0.258 | 0.245 | 0.687 | 0.207 | 0.251 | 0.311 | 0.106 | 0.167 | 0.181 | 0.105 |
| STAG-VIO (Ours) | 0.107 | 0.156 | 0.138 | 0.204 | 0.125 | 0.246 | 0.222 | 0.210 | 0.099 | 0.160 | 0.204 | 0.094 |
Across long-horizon everyday environments, STAG-VIO sustains 99.1% correct rate in market (1.276 m ATE) and 97.7% in corridor (1.376 m ATE) — the two most demanding scenes, where fixed-vocabulary baselines lose tracking as unseen or heavily occluded objects evade their segmenters. High CR paired with bounded drift is the deployment-relevant outcome: sustaining tracking through heavy dynamics, rather than terminating early.
The ablation study isolates each stage in STAG-VIO and shows that temporally stable prompts are the dominant factor behind robust dynamic-scene tracking.
The gain is a systematic shift of the entire distribution, not a single statistic: median ATE drops 76.7%, the 95th percentile falls from 2.373 m to 0.326 m, and frames exceeding 1 m drop from 16.6% to 0% — with the interquartile range tightening 7.0× and standard deviation 9.5×. Prompt stabilization doesn't just lower average error; it eliminates the catastrophic drift episodes that populate the upper tail.
Heatmap comparisons visualize where dynamic objects repeatedly disturb feature selection and optimization. STAG-VIO suppresses unstable regions while preserving reliable static constraints for visual-inertial tracking.
Densely sweeping Nmax ∈ [40, 200] and Dmin ∈ [5, 40] px, VINS-Mono spans 0.113–7.668 m with 78.9% of configurations exceeding 1 m — accuracy hinges on a well-tuned Dmin (ρ = −0.77). STAG-VIO holds a median of 0.229 m (10.2× lower) with a 25× tighter spread, essentially uncorrelated with both parameters: ANMS-based replenishment enforces spatial coverage internally, making nearly the entire parameter space deployable rather than a narrow sweet spot.
@article{zhou2026stagvio, title = {STAG-VIO: Stabilized Prompt-to-Geometry Interface for Robust Dynamic Visual-Inertial Odometry}, author = {Zhou, Rui and Liu, Jingbin and Xie, Junbin and Zhang, Jianyu and Hu, Yingze and Zhao, Jiele}, journal = {arXiv preprint}, year = {2026} }