PaperlayerView exact revision on arXiv

Source-linked AI summary

MotionGS-SLAM: Event-Modulated Gaussian Splatting for Motion-Blur Robust SLAM

Zhiqiang Hu, Shouren Huang, Masatoshi Ishikawa

arXiv:2608.15024v1cs.ROcs.AIcs.CV

TL;DR

Motion blur degrades visual SLAM tracking and reconstruction, especially during aggressive motion and long exposures. MotionGS-SLAM models blur formation with event-guided Gaussian rendering and significantly outperforms state-of-the-art methods in tracking accuracy and reconstruction quality under severe blur.

  • Problem

    Visual GS-SLAM methods rely on sharp images, but motion blur corrupts feature details and degrades camera tracking and 3D reconstruction.

  • Method

    MotionGS-SLAM uses event-guided forward blur modeling with a dual-modulation Gaussian kernel that adapts projected shape and temporal sampling to local motion.

  • Results

    MotionGS-SLAM significantly outperforms visual and GS-SLAM baselines in tracking accuracy and reconstruction quality on challenging synthetic and real-world sequences with severe motion blur.

  • Takeaways & Limitations

    The ablations support modeling both the spatial and temporal formation of blur for robust SLAM under severe motion.

Abstract

from arXiv · show

Current Vision-based SLAM systems fail catastrophically when motion blur corrupts the visual input, as they attempt the ill-posed inverse problem of recovering sharp content from degraded observations. We present MotionGS-SLAM, which fundamentally reimagines motion blur handling through a paradigm shift: rather than removing blur artifacts, we reformulate the challenge as a well-constrained forward problem that generatively models blur formation within the rendering pipeline. By leveraging event cameras' microsecond temporal resolution and immunity to motion blur, we introduce a novel event-modulated Gaussian kernel that dynamically adapts each Gaussian's rasterization based on precise motion cues. Our dual-modulation mechanism transforms 2D Gaussian projections from isotropic dots into anisotropic, motion-aligned elliptical brush strokes (spatial modulation) while adaptively varying exposure integral sampling density based on local velocity (temporal modulation). This physics-based approach enables joint optimization of intra-exposure camera trajectories and 3D scene geometry through blur-aware photometric and event-based constraints. Extensive experiments demonstrate significant improvements over state-of-the-art methods in trajectory accuracy and map quality under severe high-motion conditions.

I. INTRODUCTION

MotionGS-SLAM addresses severe motion blur in visual SLAM by modeling blur formation generatively within the rendering process rather than attempting deblurring. It uses event-guided Gaussian modulation and joint trajectory-scene optimization to improve tracking and reconstruction under challenging motion.

  • Motivation: 3D Gaussian Splatting enables dense, photorealistic scene representations in real-time, extending visual SLAM beyond its traditionally sparse geometric maps.SLAM supports robot navigation and interaction with unknown environments.
  • Challenge: Severe motion blur violates GS-SLAM’s reliance on sharp images, especially during aggressive motion or long exposures in low-light environments.Deblurring methods struggle because they invert a highly degraded signal instead of modeling the underlying physical process.
  • Approach: MotionGS-SLAM reformulates blur as the physical consequence of camera motion during exposure and models its formation generatively within the renderer using event-camera guidance.The system optimizes continuous intra-exposure camera trajectories parameterized by boundary poses and enforces photometric consistency with physically rendered blur simulations.
  • Results: The system jointly optimizes selected keyframe trajectories and the 3D Gaussian scene using photometric reconstruction and event-based geometric constraints, producing sharp, geometrically accurate maps despite degraded imagery.The paper claims significant improvements over state-of-the-art visual and GS-SLAM methods in tracking accuracy and reconstruction quality without a depth sensor.
  • Approach: The event-modulated Gaussian kernel uses local event-derived motion statistics to jointly adapt projected Gaussian shape and temporal sampling density through dual modulation.A 4D spatio-temporal hash grid efficiently supports event association and motion-statistics extraction.

II. RELATED WORKS · A. Gaussian Splatting based SLAM · B. Event-based 3D Gaussian Splatting/NeRF

This related-work section traces Gaussian Splatting-based SLAM from real-time dense mapping and unified tracking-reconstruction systems to efforts addressing motion blur, then introduces event-camera integration with NeRF/3DGS. It emphasizes both the viability of explicit differentiable representations and the limitations of image-only blur recovery.

  • A. Gaussian Splatting based SLAM: 3D Gaussian Splatting enabled SLAM systems that create dense, photorealistic maps in real time.
  • A. Gaussian Splatting based SLAM: GS-SLAM, SplaTAM, MonoGS, Sgs-slam, Large spatial model, and Compact 3d gaussian splatting integrated 3DGS into unified tracking and mapping frameworks.
  • A. Gaussian Splatting based SLAM: These systems demonstrated the viability of explicit, differentiable representations for localization and scene reconstruction.
  • A. Gaussian Splatting based SLAM: Gaussian Splatting-based SLAM methods remain brittle because they assume sharp, well-exposed images.
  • A. Gaussian Splatting based SLAM: Image-only approaches such as I2-SLAM model camera imaging but attempt to recover information from corrupted signals, failing when severe motion makes blur too ambiguous to invert reliably.
  • B. Event-based 3D Gaussian Splatting/NeRF: Event cameras support three-dimensional scene reconstruction through asynchronous operation and inherent immunity to motion blur.
  • B. Event-based 3D Gaussian Splatting/NeRF: Recent work combines event streams with NeRF/3DGS, including E2NeRF’s blurred-image supervision and event-based constraints for high-fidelity NeRF reconstruction.
  • B. Event-based 3D Gaussian Splatting/NeRF: Ev-DeblurNeRF extends this direction with a learned event-to-intensity mapping to suppress degradation.

III. METHOD · A. Motion Blur Image Formation Model

The method represents scenes with 3D Gaussians and forms motion-blurred images by integrating sharp renderings along a continuously moving camera trajectory. This trajectory is parameterized by exposure-start and exposure-end poses and approximated using discrete latent images.

  • III. METHOD: MotionGS-SLAM’s core architecture is presented in Fig. 2.
  • A. Motion Blur Image Formation Model: The scene is represented as 3D Gaussians, each defined by a mean, covariance, color, and opacity.The Gaussian set is denoted G = {g_i}.
  • A. Motion Blur Image Formation Model: In static rendering, each 3D Gaussian is projected onto the image plane using a single camera pose.
  • A. Motion Blur Image Formation Model: The final pixel color C(u) is computed through α-blending.
  • A. Motion Blur Image Formation Model: Motion blur is modeled as light integration over a non-zero exposure interval while the camera follows a continuous trajectory T(t).The trajectory spans exposure start and end times, and the blurred image is the temporal integral of infinitesimally short exposures.
  • A. Motion Blur Image Formation Model: The blurred image is approximated by discretely summing N latent sharp images rendered at intermediate camera poses.Each sharp image is rendered as C(T(t_k), u).
  • A. Motion Blur Image Formation Model: Camera motion is parameterized by exposure-start pose T_st and exposure-end pose T_ed in SE(3).The pose at any exposure time is defined using exponential and logarithm maps on SE(3).

B. Motion Blur Aware Tracker

The motion-blur-aware tracker associates each visible Gaussian with relevant exposure-window events, extracts local motion statistics, and uses them to modulate spatial kernels and temporal sampling. It jointly estimates exposure start and end poses through physics-based photometric and event constraints while preserving the underlying 3D geometry.

  • Real-Time Event-Gaussian Association: The hash-based association reduces complexity from O(Ngauss×Nevent) to O(Ngauss×K), where K is the small average number of events returned per Gaussian.This avoids the computational infeasibility of brute-force search for real-time operation.
  • Real-Time Event-Gaussian Association: A 4D multi-resolution hash grid efficiently indexes events by spatial coordinates, timestamp, and polarity for real-time Gaussian–event association.Each event is represented as e = (x, y, t, p), with polarity indicating brightness decrease or increase.
  • Real-Time Event-Gaussian Association: The tracker queries events within each projected Gaussian’s 3σ image-space bounding box and exposure window, then filters candidates using a Mahalanobis-distance criterion.The threshold is typically τ ≈9, corresponding to the 99% confidence contour of a 2D Gaussian.
  • Event-Modulated Gaussian Kernel: Polarity-weighted event displacements yield each Gaussian’s dominant local motion flow, whose magnitude and angle determine velocity and direction for dual spatial-temporal kernel modulation.Opposite-polarity events moving consistently reinforce one another, while inconsistent noise cancels out.
  • Blur Aware Tracking Optimization: The tracker estimates exposure endpoint poses Tst and Ted by minimizing an objective combining physics-based photometric consistency with event terms.The rendered blurred image is required to align with the image captured by the camera sensor, while λevt balances the event terms.

1) Keyframe Selection: · 2) Bundle Adjustment:

Keyframes are selected using view-overlap and translation thresholds, with redundancy removed when the keyframe limit is reached. Bundle adjustment aligns rendered intensity changes with observed event polarities, providing geometric constraints in ambiguous or low-texture regions.

  • 1) Keyframe Selection:: A new keyframe is inserted when visible-Gaussian IoU with the latest keyframe falls below τoverlap.The overlap is measured using the visible Gaussians.
  • 1) Keyframe Selection:: A new keyframe is also inserted when camera translation exceeds τtrans · ¯d, where ¯d denotes average scene depth.This criterion adapts the translation threshold to scene scale through average depth.
  • 1) Keyframe Selection:: When the keyframe limit is reached, the keyframe with highest neighbor overlap is removed to preserve view diversity and bound computation.Removal targets the most redundant keyframe according to neighbor overlap.
  • 2) Bundle Adjustment:: Bundle adjustment uses an event alignment loss to match rendered intensity changes with brightness changes observed by the event camera.The estimated camera motion must reproduce the observed brightness changes in the rendered 3D scene.
  • 2) Bundle Adjustment:: The loss enforces that each event polarity pi ∈ {+1, −1} matches the sign of the corresponding logarithmic intensity change.Deviations from the event generation model are penalized.
  • 2) Bundle Adjustment:: Rendered intensity I(t) is evaluated at event timestamp ti with temporal offset δ and contrast threshold θ, yielding strong constraints in ambiguous or low-texture regions.These variables define the event-based constraint used for geometric supervision.

C. Mapping with Keyframe Management

The mapping back-end maintains long-term consistency by jointly refining 3D Gaussians and keyframe poses within a sliding window. Its blur-consistent objective reproduces blurry images and event streams while imposing motion-aligned projected Gaussian shapes.

  • Mapping back-end: The mapping back-end uses a sliding window of keyframes to jointly refine 3D Gaussians and camera poses for long-term map consistency.This refinement maintains physical agreement between sensor observations over time.
  • Bundle adjustment: Bundle adjustment jointly optimizes the Gaussian map G and keyframe poses {T} using a comprehensive, blur-consistent objective.The optimization is defined over the current keyframe window W.
  • Data consistency: The data-consistency terms require the optimized map and poses to reproduce each keyframe’s observed blurry images and event streams.This enforces consistency across time.
  • Shape prior: The image-plane shape prior encourages projected 3D Gaussians to become gently elongated along detected motion, while remaining isotropic without detected motion.The prior changes the projected 2D kernel without altering the underlying 3D geometry.

IV. EXPERIMENTS

The experiments rigorously evaluate MotionGS-SLAM to validate its effectiveness.

  • The section conducts a series of experiments to evaluate MotionGS-SLAM.
  • The experiments are designed as rigorous validation of MotionGS-SLAM’s effectiveness.
  • The section’s stated objective is to validate the effectiveness of MotionGS-SLAM experimentally.

A. Datasets and Metrics

The evaluation uses synthetic and real-world sequences, including a motion-blur- and noise-challenging EventReplica adaptation and three Color-DAVIS346 indoor/outdoor scenes. Trajectory accuracy is measured with ATE, while synthetic mapping quality is assessed using PSNR, SSIM, and LPIPS.

  • Datasets: The method is evaluated on both synthetic and real-world sequences for comprehensive analysis.
  • Datasets: EventReplica adapts Replica by averaging many rendered intermediate frames to produce physically realistic motion blur, with Poisson-Gaussian noise added for low-light conditions.The synthetic dataset provides ground-truth sharp images for objective reconstruction evaluation.
  • Datasets: The real-world dataset contains 3 indoor/outdoor scenes recorded with a Color-DAVIS346 camera capturing 346×260 color frames and color events under low-light handheld motion.Each scene includes rich textures and color information and is recorded across multiple viewpoints with varying motion blur.
  • Metrics: Trajectory accuracy is evaluated with Absolute Trajectory Error (ATE) [cm], while synthetic mapping quality uses PSNR, SSIM, and LPIPS [24].

B. Implementation Details · C. Comparison with State-of-the-Art · D. Ablation Study

MotionGS-SLAM is implemented with specified event-kernel and loss settings and evaluated against representative SLAM baselines for reconstruction and tracking under severe blur. Ablations show that event integration and joint spatial-temporal modulation are central to achieving the strongest reported performance.

  • B. Implementation Details: The system runs in PyTorch on a single NVIDIA RTX 4080 GPU, builds on MonoGS, and uses n0 = 9, β = 0.05, λevt = 2.0, λphoto = 1.0, and λshape = 0.2.Competing methods use their official recommended settings for fair evaluation.
  • C. Comparison with State-of-the-Art: MotionGS-SLAM is compared with PhotoSLAM, MonoGS, and EGS-SLAM on EventReplica for reconstruction fidelity and tracking accuracy under severe motion blur.Evaluation includes low-light and long-exposure settings, with PSNR, SSIM, LPIPS, rendering FPS, and tracking accuracy considered.
  • C. Comparison with State-of-the-Art: On the real-world dataset, MotionGS-SLAM is the only method described as reconstructing sharp, geometrically consistent results, while image-only baselines show ghosting and EGS-SLAM retains soft details.The qualitative comparison uses challenging low-light handheld sequences.
  • D. Ablation Study: The ablation study evaluates five variants, from blurry-image-only MonoGS and a blur-model-only system through event association, spatial-only, temporal-only, and full dual-modulation configurations.A5 enables both spatial and temporal modulation, whereas A3 and A4 isolate the two kernel components.
  • D. Ablation Study: 6.10cm ATE on room0 is achieved by event integration, reducing the baseline A0 error by 52%, whereas the blur model alone reduces ATE from 12.76cm to 11.80cm.These results establish the critical role of event data for motion-blur handling.
  • D. Ablation Study: Spatial modulation improves room0 LPIPS from 0.340 to 0.250, described as the largest perceptual gain among the evaluated components.The result supports event-guided anisotropic deformation for modeling motion streaks.
  • D. Ablation Study: The qualitative ablation shows progression from blurry A2 reconstructions to sharper A3 results and high-fidelity reconstruction with A5.The comparison highlights reconstructed details under severely blurred input.
  • D. Ablation Study: The full A5 model achieves the best reported metrics—ATE: 3.98cm, PSNR: 24.55dB, and LPIPS: 0.214—by combining spatial and temporal modulation.The combined model surpasses either modulation component alone while maintaining computational efficiency.

V. SUMMARY

MotionGS-SLAM addresses high-motion SLAM by modeling motion blur as a physical process within rendering rather than correcting it as an artifact. Its central contribution is an event-modulated Gaussian kernel guided by high-frequency event data.

  • System overview: MotionGS-SLAM is a robust SLAM system designed for high-motion scenarios.The passage states that the system excels under high motion.
  • System overview: The method reformulates motion blur as a physical process generatively modeled within the rendering pipeline.This shifts blur handling from artifact correction to physics-based synthesis.
  • Core contribution: Its central contribution is an event-modulated Gaussian kernel that uses high-frequency event data to guide physics-based synthesis.The kernel provides event-guided modulation for the rendering process.
Loading 2608.15024v1…