VIGS-SLAM and VINGS-Mono

Related docs: VINS-Mono and VINS-Fusion, OpenVINS, GS-SLAM and MonoGS, S3PO-GS, and Gaussian Splatting for Driving.

Executive Summary

VINGS-Mono and VIGS-SLAM are visual-inertial Gaussian SLAM systems from the 2025-2026 wave. They address a clear weakness of pure visual Gaussian SLAM: photometric tracking and Gaussian map optimization degrade under blur, low texture, exposure changes, and large-scale outdoor motion.

VINGS-Mono targets monocular large-scene Gaussian mapping with a VIO front end, 2D Gaussian map, novel-view-synthesis loop closure, and Dynamic Eraser. It reports kilometer-scale outdoor capability and a smartphone-style camera plus low-frequency IMU use case.

VIGS-SLAM tightens the visual-inertial formulation by jointly optimizing camera poses, depths, and IMU states in a unified framework, with robust IMU initialization, time-varying bias modeling, loop closure, and consistent Gaussian updates. For AVs, this lineage is more relevant than camera-only Gaussian SLAM, but it is still not a production-grade localization stack.

Core Idea

The common idea is to let inertial sensing stabilize Gaussian visual SLAM.

Core elements:

• Monocular RGB supplies appearance and visual geometry.
• IMU supplies high-rate motion and gravity constraints.
• Gaussian maps provide dense, renderable scene representation.
• VIO front ends provide metric or scale-stabilized trajectories.
• Loop closure corrects trajectory and Gaussian map consistency.
• Dynamic-object handling tries to prevent transient objects from polluting maps.

VINGS-Mono emphasizes large-scale system design. VIGS-SLAM emphasizes tighter visual-inertial optimization and IMU-state modeling.

Pipeline

VINGS-Mono:

1. Process RGB frames and low-rate IMU data in a VIO front end.
2. Use dense bundle adjustment and uncertainty estimation to recover geometry and pose.
3. Incrementally build a 2D Gaussian map.
4. Use a sample-based rasterizer, score manager, and pose refinement for efficient mapping.
5. Use NVS loop closure to detect and correct loops through Gaussian rendering.
6. Use Dynamic Eraser to reduce dynamic-object artifacts.

VIGS-SLAM:

1. Initialize visual-inertial state robustly.
2. Track camera poses with visual and inertial constraints.
3. Jointly refine camera poses, depths, and IMU states.
4. Model time-varying IMU biases.
5. Build and update a 3D Gaussian map.
6. Apply loop closure with consistent Gaussian map updates.
7. Render dense maps for reconstruction and evaluation.

Strengths

• IMU improves robustness over purely visual Gaussian SLAM.
• Gravity and inertial propagation help with blur, low texture, and fast motion.
• VINGS-Mono addresses large outdoor scenes and reports up to tens of millions of Gaussian ellipsoids.
• NVS-based loop closure is a novel use of Gaussian rendering.
• Dynamic Eraser acknowledges real outdoor dynamic content.
• VIGS-SLAM models IMU initialization and time-varying bias explicitly.
• Both methods are closer to practical robotics than RGB-only indoor Gaussian SLAM.

Limitations

• Monocular visual-inertial scale and bias observability still require good excitation and calibration.
• Camera appearance remains fragile under weather, night, glare, lens dirt, and low texture.
• Low-frequency phone IMUs are not equivalent to automotive-grade synchronized IMUs.
• Dynamic-object modules can miss stationary or slow-moving temporary objects.
• NVS loop closure can be confused by repeated gates, corridors, road segments, or stands.
• Gaussian map quality and pose quality must be evaluated separately.
• No native LiDAR, radar, wheel odometry, GNSS, or HD-map constraints in the core descriptions.
• Production covariance, integrity, and failover behavior remain open.

AV Relevance

This lineage is relevant to AVs because visual-inertial Gaussian SLAM is the natural bridge between camera-only Gaussian maps and classical VIO.

Potential roles:

• Dense visual map construction with IMU-stabilized poses.
• Visual-inertial benchmark against OpenVINS, VINS-Fusion, ORB-SLAM3, DROID-SLAM, and MASt3R-SLAM.
• Offline reconstruction for depots, terminals, service roads, and structured urban scenes.
• Research into Gaussian-rendering loop closure and visual map QA.

Production use still needs multi-sensor fusion. A road or airside AV should combine VIO with wheel odometry, LiDAR/radar, GNSS/RTK when valid, map constraints, dynamic-object policies, and safety monitors.

Indoor/Outdoor Notes

Indoor: Visual-inertial Gaussian SLAM is useful for corridors, warehouses, terminals, hangars, labs, and AR-like mapping, especially where IMU helps through texture-poor segments.

Outdoor: VINGS-Mono is explicitly aimed at large outdoor scenes. VIGS-SLAM targets challenging visual-inertial conditions. Outdoor success still depends on lighting, texture, calibration, and dynamic-object control.

Airside: Potentially useful for offline mapping and visual inspection around terminals and service roads. For open aprons, repeated stands, wet pavement, aircraft motion, and night floodlights make standalone visual-inertial Gaussian localization risky.

Comparison

Method	Sensors	Main distinction	AV interpretation
VINGS-Mono	Monocular RGB + optional/low-rate IMU	Large-scene 2D Gaussian map, NVS loop closure, Dynamic Eraser	Large-scale visual-inertial Gaussian mapping baseline
VIGS-SLAM	RGB + IMU	Tight VI optimization of poses, depths, IMU states, loop-consistent Gaussians	More estimator-centered VI Gaussian SLAM
S3PO-GS	Monocular RGB	Scale-consistent pointmap Gaussian outdoor SLAM	Foundation geometry without IMU
ORB-SLAM3/OpenVINS	Camera + IMU	Mature sparse VIO/SLAM	Strong classical baselines
Gaussian-LIC2	LiDAR + IMU + camera	Metric LIC Gaussian map	More AV-practical sensor mix

Evaluation

Key metrics:

• ATE/RPE with explicit alignment.
• Scale drift and yaw drift over long routes.
• IMU bias stability and initialization success.
• Loop-closure precision, recall, and correction magnitude.
• Rendering metrics: PSNR, SSIM, LPIPS.
• Map size, primitive count, runtime, and GPU memory.
• Dynamic-object ghost rate.
• Failure count under blur, exposure change, low texture, and aggressive motion.

For AV/airside use, add drift per kilometer, lateral/yaw error at operational waypoints, repeated-route consistency, false loop closures in repeated infrastructure, and performance under night/rain/glare.

Implementation Notes

• Use hardware-synchronized camera and IMU logs for serious evaluation.
• Calibrate camera intrinsics, camera-IMU extrinsics, and temporal offset.
• Record IMU rate and grade; do not generalize phone-IMU results to vehicle IMUs or the reverse.
• Keep VIO health outputs separate from rendering metrics.
• Validate NVS loop closures against independent pose truth before applying them to maps.
• Maintain a dynamic-object map policy for parked vehicles, aircraft, GSE, and temporary barriers.
• Pin CUDA, PyTorch, and rasterizer dependencies if using released research code.

Practical Recommendation

Use VINGS-Mono and VIGS-SLAM to understand the visual-inertial Gaussian SLAM direction. For AV and airside systems, benchmark them offline against classical VIO and multi-sensor localization. Do not use them as standalone operational localization until metric integrity, loop-closure safety, dynamic handling, and adverse-condition behavior are independently validated.

Sources

• Wu, Zhang, Tie, Ai, Gan, and Ding, "VINGS-Mono: Visual-Inertial Gaussian Splatting Monocular SLAM in Large Scenes." https://arxiv.org/abs/2501.08286
• Zhu, Zhang, Li, Haala, Pollefeys, and Barath, "VIGS-SLAM: Visual Inertial Gaussian Splatting SLAM." https://arxiv.org/abs/2512.02293
• VIGS-SLAM project page. https://vigs-slam.github.io/
• Local context: VINS-Mono and VINS-Fusion
• Local context: OpenVINS
• Local context: Gaussian Splatting for Driving