GaussianFlow SLAM

Video

Abstract

Gaussian splatting has recently gained traction as a compelling map representation for SLAM systems, enabling dense and photo-realistic scene modeling. However, its application to monocular SLAM remains challenging due to the lack of reliable geometric cues from monocular input. Without geometric supervision, mapping or tracking could fall in local-minima, resulting in structural degeneracies and inaccuracies.

To address this challenge, we propose GaussianFlow SLAM, a monocular 3DGS-SLAM that leverages optical flow as a geometry-aware cue to guide the optimization of both the scene structure and camera poses. By encouraging the projected motion of Gaussians, termed GaussianFlow, to align with the optical flow, our method introduces consistent structural cues to regularize both map reconstruction and pose estimation. Furthermore, we introduce normalized error-based densification and pruning modules to refine inactive and unstable Gaussians, thereby contributing to improved map quality and pose accuracy.

Experiments conducted on public datasets demonstrate that our method achieves superior rendering quality and tracking accuracy compared with state-of-the-art algorithms.

Method

GaussianFlow SLAM is a monocular 3D Gaussian Splatting SLAM framework that uses dense optical flow as a geometry-aware cue for optimizing the 3DGS map and camera poses. Given incoming monocular images, the system alternates between tracking and mapping while maintaining a shared 3DGS map. By aligning GaussianFlow with dense optical flow, the system provides structural supervision even without depth measurements. This allows the Gaussians to move toward geometrically consistent positions while also providing feedback for pose estimation.

Our pipeline consists of three recurrent optimization modules: initial pose optimization, GaussianFlow-guided dense bundle adjustment, and multi-view Gaussian optimization. In each recurrent update, GaussianFlow rendered from the current 3DGS map is fed into a ConvGRU-based optical flow module, which predicts refined optical flow. The resulting flow is then used by optimization kernels to update either the camera poses or the 3DGS map. The optical flow and GaussianFlow progressively converge toward geometrically consistent correspondences.

To further improve map quality, we introduce normalized error-based densification and pruning. These modules selectively split under-reconstructed Gaussians and remove unstable floaters using per-Gaussian errors. This helps produce more accurate geometry, cleaner reconstructions, and better pose estimation.

Demo Results

EuRoC — MH01

TUM RGB-D — fr1/desk

EuRoC — V101

EuRoC — MH05

BibTeX

@article{seo2026gaussianflowslam,
  title     = {GaussianFlow SLAM: Monocular Gaussian Splatting SLAM Guided by GaussianFlow},
  author    = {Seo, Dong-Uk and Jeon, Jinwoo and Lee, Eungchang Mason and Myung, Hyun},
  journal   = {IEEE Robotics and Automation Letters},
  year      = {2026}
}