SCE-SLAM: Scale-Consistent Monocular SLAM via Scene Coordinate Embeddings

TL;DR

Monocular SLAM suffers from scale drift due to lack of global scale constraints across temporal windows. SCE-SLAM prevents scale fragmentation by learning patch-level scene coordinate embeddings that encode geometry under a canonical scale and by propagating this information with geometry-guided attention, followed by scene coordinate bundle adjustment that anchors current estimates to the reference scale. The system combines a flow-based local constraint with a scale-aware scene-coordinate branch in a dual-branch architecture, and employs a two-stage bootstrapping BA to enforce long-term scale consistency. Experiments on KITTI, Waymo, and Virtual KITTI show substantial ATE improvements, real-time performance, and robust scale consistency across large-scale sequences, outperforming prior frame-to-frame and some global approaches. The work offers practical impact for scalable, resource-efficient monocular SLAM in autonomous navigation and 3D reconstruction tasks.

Abstract

Monocular visual SLAM enables 3D reconstruction from internet video and autonomous navigation on resource-constrained platforms, yet suffers from scale drift, i.e., the gradual divergence of estimated scale over long sequences. Existing frame-to-frame methods achieve real-time performance through local optimization but accumulate scale drift due to the lack of global constraints among independent windows. To address this, we propose SCE-SLAM, an end-to-end SLAM system that maintains scale consistency through scene coordinate embeddings, which are learned patch-level representations encoding 3D geometric relationships under a canonical scale reference. The framework consists of two key modules: geometry-guided aggregation that leverages 3D spatial proximity to propagate scale information from historical observations through geometry-modulated attention, and scene coordinate bundle adjustment that anchors current estimates to the reference scale through explicit 3D coordinate constraints decoded from the scene coordinate embeddings. Experiments on KITTI, Waymo, and vKITTI demonstrate substantial improvements: our method reduces absolute trajectory error by 8.36m on KITTI compared to the best prior approach, while maintaining 36 FPS and achieving scale consistency across large-scale scenes.

Figures (7)

• Figure 1: Kilometer-Scale Sequences Performance. Remarkably relieving scale drift problem, our method, even without loop closure, achieves superior accuracy and robustness with high efficiency, obtaining the best ATE in average and std dev. across 11 KITTI geiger2012kitti scenes. Marker size denotes runtime on sequence 05.
• Figure 2: Overview of SCE-SLAM. Given patch features from a DINOv3-augmented backbone, we construct a patch graph distinguishing active patches (current window) from reference patches (scale anchors). Geometry-Guided Scale Propagation operates through dual branches: the flow branch initializes correspondences and scale reference; the scene coordinate branch maintains embeddings $\mathbf{h}^{\text{xyz}}$, aggregates scale from reference patches via geometry-modulated attention to compute spatial correlation $\mathbf{f}_k^{\text{sc}}$, and predicts metric increments $\Delta \mathbf{X}_k$. Scene Coordinate Bundle Adjustment jointly refines poses and depths using both predictions for scale-consistent SLAM.
• Figure 3: Visual Odometry Comparison on the KITTI Datasets w/o Loop Closure. We visualize the estimated trajectory with the scale from the first 20 timestamps. The results show the strong VO capability of our method. Especially, we visualize the estimated relative scale by color (left), and logarithmic bias per time (right). The reduced scale drift demonstrates our advantage in the global scale consistency.
• Figure 4: Trajectory on 4Seasons Neighborhood. Blue: our prediction; red dashed: ground truth. Our method successfully closes the loop with scale consistency, while DPV-SLAM++ fails.
• Figure 5: Multi-view patch sampling consistency. Patch tracking across three consecutive frames. Blue boxes show successful tracks across views, red boxes show tracking failures.