MyGO-Splat: Multi-Objective Closed-Loop Geometric Feedback for RGB-Only Gaussian SLAM

Abstract: Real-time monocular Simultaneous Localization and Mapping (SLAM) fundamentally suffers from scale ambiguity and a lack of geometric self-correction. While 3D Gaussian Splatting (3DGS) enables high-fidelity rendering, existing RGB-only systems remain open-loop because depth priors are injected into mapping but refined geometry cannot effectively regulate tracking drift. We present MyGO-Splat, a closed-loop Gaussian SLAM framework that analytically rasterizes Gaussian primitives into pixel-wise depth and surface normals, allowing the map to actively supervise camera pose optimization. To bridge monocular priors and scale consistency, our framework introduces scale-aware adaptive alignment that projects foundation-model depth estimates into the globally optimized Gaussian space, forming a self-correcting cycle for scale feedback. Extensive evaluations show that this closed-loop design improves scale stability and appearance-geometry consistency, achieving performance comparable to RGB-D methods while using only monocular input.

• The paper introduces a closed-loop geometric feedback scheme that integrates refined Gaussian maps with monocular pose estimation to self-correct scale drift.
• It employs analytical rasterization for precise pixel-level depth and normal computation, enhancing both tracking and mapping performance.
• Experimental results on Replica, TUM, and ScanNet demonstrate competitive pose accuracy, superior rendering quality, and robust mesh reconstruction compared to RGB-D methods.

Multi-Objective Closed-Loop Geometric Feedback for RGB-Only Gaussian SLAM: MyGO-Splat

Motivation and Problem Formulation

Monocular SLAM systems have historically struggled with scale ambiguity and limited geometric self-correction due to the absence of metric depth information. The decoupled nature of tracking and mapping in conventional RGB-only pipelines typically results in persistent scale drift and fragile long-term geometric consistency. While 3D Gaussian Splatting (3DGS) enhances real-time photorealistic map expressiveness, existing systems largely treat depth priors as auxiliary supervision, failing to integrate refined geometric representations back into the pose optimization process. This paper introduces "MyGO-Splat: Multi-Objective Closed-Loop Geometric Feedback for RGB-Only Gaussian SLAM" (2606.29738), which systematically upgrades conventional open-loop architectures into a closed-loop system where the Gaussian map actively supervises the tracking frontend, enforcing multi-view consistency and robust scale regulation.

Methodology

Closed-Loop Geometric Feedback Architecture

MyGO-Splat is formulated as a closed-loop framework comprising three synergistic modules:

1. Scale-Regulated Online Tracking: A recurrent optical flow-based pose estimation approach constructs a keyframe graph, leveraging a learning-based dense flow network for robust camera pose and local geometry estimation. Keyframe selection uses flow-based criteria, and bundle adjustment (BA) is performed on keyframes.
2. Closed-Loop Geometric Feedback: The system periodically projects foundation-model depth estimates (e.g., from VGGT) into the globally optimized Gaussian space via uncertainty-weighted metric synchronization. Optimal scale and shift parameters are computed to align monocular priors with the geometric space maintained by SLAM, forming a self-correcting cycle for depth feedback.
3. Analytical Rasterization of Gaussian Primitives: The mapping module analytically rasterizes Gaussian primitives for pixel-wise depth and normal computation, shifting from center-projection to a differentiable ray--Gaussian intersection approach. This enables active supervision with multi-view consistent geometric fields.

Joint optimization is achieved using a geometric-enhanced multi-objective loss that balances rendered appearance, depth alignment, surface continuity, and normal consistency. Analytical rasterization allows for precise geometric regularization at pixel-level granularity.

Analytical Depth and Normal Computation

Depth per pixel is computed by rasterizing the Gaussian primitives under a local affine transformation. Intersection points between camera rays and Gaussians are solved analytically, leveraging coplanar intersection properties in ray space. Surface normals are derived from the plane equations formed in ray space and transformed back to camera coordinates via the affine Jacobian. These analytical computations provide supervisory signals for optimization in both photometric and geometric domains.

Global Consistency via Loop Closure

Visual similarity-based loop closure is employed for global consistency. Keyframe features are indexed in a FAISS database, and candidates are retrieved for loop constraints. Upon successful loop detection, global BA optimizes poses and depths, eliminating cumulative drift.

Experimental Results

Extensive evaluations were conducted on Replica (synthetic), TUM RGB-D, and ScanNet datasets. MyGO-Splat executes online with performance metrics reported on pose accuracy (ATE RMSE), appearance (PSNR, SSIM, LPIPS), and geometric fidelity (accuracy, completion).

• Pose Estimation: MyGO-Splat delivers competitive ATE RMSE in RGB-only settings, approaching RGB-D methods. On challenging scenes (ScanNet), feedback from rendered geometry mitigates scale drift.
• Rendering Quality: The method achieves highest PSNR and SSIM amongst RGB-only competitors, outperforming several RGB-D baselines. Sharp textures and high-frequency details are reliably reconstructed.
• Geometric Accuracy: Mesh reconstructions extracted via marching tetrahedra and TSDF fusion exhibit high surface accuracy and completion. Analytical rasterization produces continuous, artifact-free surfaces.
• Ablation Study: Removal of loop closure, closed-loop geometric feedback, or geometric-enhanced multi-objective optimization significantly degrades both localization and geometric metrics. Closed-loop feedback suppresses floaters and ghosting, while geometric regularization ensures topological surface integrity.

Contradictory and Strong Claims

• The paper asserts that closed-loop geometric feedback in RGB-only SLAM can achieve appearance-geometry consistency equivalent to RGB-D methods without external sensors.
• It claims that analytical depth and normal computation from Gaussian parameters fundamentally upgrades the Gaussian map from a passive renderer to an optimization-aware scene representation, impacting both tracking and mapping.

Theoretical and Practical Implications

The closed-loop architecture explicitly closes the feedback loop between mapping and tracking, enabling SLAM systems to self-correct scale using only monocular input. Analytical rasterization obviates the need for depth sensors while maintaining the fidelity and continuity demanded for mesh extraction and further downstream tasks. The method integrates foundation-model geometric priors into metric space, facilitating robust operation on unconstrained sequences and paving the way for more generalizable SLAM approaches.

Future Directions

This framework is readily extensible to semantic and dynamic scene modeling. Integration of more advanced monocular geometric priors could further improve scale correction. Efficient multi-modal feedback architectures could be generalized to other neural scene representations, enabling robust SLAM in fully sensorless environments and for real-time embodied intelligence. Analytical feedback mechanisms also offer promise for hierarchical scene understanding.

Conclusion

MyGO-Splat defines a unified closed-loop geometric feedback paradigm for RGB-only Gaussian SLAM. By integrating analytical rasterization, scale-aware feedback, and multi-objective optimization, the system achieves robust, appearance-geometry consistent reconstruction and pose estimation. Quantitative and qualitative evaluations demonstrate superiority over prior RGB-only approaches and parity with RGB-D methods, establishing MyGO-Splat as a reference for monocular SLAM in the absence of depth sensors (2606.29738).