Research map / Architectures and efficiency / RNNs and normalization
SchurVINS: Schur Complement-Based Lightweight Visual Inertial Navigation System
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SchurVINS, a filter-based (EKF) visual-inertial odometry system that builds a full residual model (gradient, Hessian, covariance) and applies the Schur complement to split it into pose and per-landmark EKF updates, achieving accuracy near optimization-based methods at low CPU cost.
Problem
Visual inertial navigation needs both high accuracy and low computational cost on resource-constrained devices; filter-based VINS are efficient but less accurate, while optimization-based VINS are accurate but expensive.
Approach
Project the stacked reprojection residual into the Jacobian space to get an equivalent gradient/Hessian/covariance residual, then apply the Schur complement to separate an ego-motion EKF update from many independent single-landmark EKF updates, all inside an EKF framework built on SVO2.0.
Key results
- On EuRoC, SchurVINS achieves 0.075 m average APE RMSE, the lowest among filter-based methods (OpenVINS 0.096), and comparable to optimization-based BASALT (0.072) while DM-VIO reaches 0.064.
- On TUM-VI, SchurVINS averages 0.182 m RMSE versus 0.314 for OpenVINS (4-frame window setting) and 0.141 for DM-VIO.
- CPU usage averages 18% versus 37% for OpenVINS and 89% for SVO2.0 on EuRoC; total time 3.83 ms per frame on MH01 versus 8.53 ms for OpenVINS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SchurVINS instead of DM-VIO | VINS on EuRoC · EuRoC | APE RMSE: 0.064 -> 0.075 | Hurt |
| Use SchurVINS instead of DM-VIO | VINS on TUM-VI · TUM-VI | APE RMSE: 0.141 -> 0.182 | Hurt |
| Use SchurVINS instead of OpenVINS | VINS on EuRoC · EuRoC | Avg CPU: 37 -> 18 | Helped |
| Use SchurVINS instead of OpenVINS | VINS on EuRoC MH01 · EuRoC MH01 | Total time (ms): 10.91 -> 3.83 | Helped |
| Add EKF-based landmark solver (ablation) | SchurVINS · EuRoC | APE RMSE: 0.107 -> 0.075 | Helped |
| Add EKF-based landmark solver instead of Gauss-Newton landmark estimation (ablation) | SchurVINS · EuRoC | APE RMSE: 0.073 -> 0.075 | Hurt |
| Use EKF-based landmark estimation instead of Gauss-Newton landmark estimation (ablation) | SchurVINS · EuRoC MH01 | optimizeStructure time (ms): 0.42 -> 0.11 | Helped |
Surprising findings
- The EKF-based landmark solver is slightly less accurate than the Gauss-Newton landmark solver of SVO2.0 but is about 3x faster in the structure-optimization step.
Limitations
- Landmark estimation uses only observations in the sliding window, causing slight accuracy loss relative to Gauss-Newton landmark optimization.
- Lacks global bundle adjustment, loop closure and pose graph optimization, so accuracy trails the best optimization-based methods on some sequences (e.g., DM-VIO on TUM-VI).
Experimental setup
Datasets: EuRoC MAV TUM-VI
Benchmarks: EuRoC MAV TUM-VI
Metrics: APE RMSE CPU usage run time (ms)
Baselines: OKVIS VINS-Mono Kimera ICE-BA SVO2.0 BASALT DM-VIO MSCKF ROVIO OpenVINS SMSCKF
Builds on: SVO2.0 MSCKF OpenVINS Schur complement
Closest papers by meaning
- EqNIO: Subequivariant Neural Inertial Odometry (ICLR 2025 · similarity 0.82)
- Adaptive VIO: Deep Visual-Inertial Odometry with Online Continual Learning (CVPR 2024 · similarity 0.80)
- iMoT: Inertial Motion Transformer for Inertial Navigation (AAAI 2025 · similarity 0.80)
- Weather-Robust Cross-View Geo-Localization via Prototype-Based Semantic Part Discovery (NeurIPS 2026 · similarity 0.79)
- Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics (CVPR 2025 · similarity 0.79)
- OVER-NAV: Elevating Iterative Vision-and-Language Navigation with Open-Vocabulary Detection and StructurEd Representation (CVPR 2024 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.