Introduction
Simultaneous Localization and Mapping (SLAM) is a cornerstone of robotics, yet standard systems often fail when faced with "messy" real-world data. Disjoint video sessions—caused by rapid camera turns, occlusions, or multiple agents mapping the same space—present a significant challenge for trajectory alignment. Traditional methods rely on hand-crafted features that struggle with blur or low texture, while modern deep-learning approaches often assume small-baseline continuity. This paper introduces a robust system for Multi-Session SLAM that bridges this gap by coupling deep optical flow with a novel differentiable solver for wide-baseline pose estimation.
Proposed Methodology
The system utilizes a unified backbone architecture inspired by iterative flow networks like RAFT. Unlike traditional SLAM, which separates tracking and mapping, this approach treats matching as an iterative optimization problem handled by a Recurrent Neural Network (RNN).
The architecture iteratively refines optical flow and camera poses. For standard tracking (Visual Odometry), the system employs Bundle Adjustment. However, for the more difficult task of connecting distant, disjoint sessions (Trajectory Alignment), it introduces a specialized geometric solver.
Differentiable SED Solver
The "secret sauce" of the proposed framework is the differentiable solver layer that minimizes the Symmetric Epipolar Distance (SED). This allows the entire network to be trained end-to-end, enabling the model to learn which features are most reliable for pose estimation.
To handle the non-convex nature of the SED objective, the authors implement a two-stage process:
- Pre-conditioning: A weighted, dense version of the classic 8-point algorithm provides a robust initial guess.
- SED Optimization: The differentiable solver polishes this guess to high precision.
This combination allows the system to handle both small and wide baselines effectively. Furthermore, the system estimates a 7-degree-of-freedom Sim(3) transformation, which is essential for aligning monocular sessions where the absolute scale may vary between independent video clips.
Experimental Evaluation
The model was trained on diverse datasets, including Scannet (indoor), Megadepth (outdoor), and TartanAir (synthetic). Evaluation focused on both isolated two-view pose estimation and full multi-session SLAM performance.
Key Results
- Superior Accuracy: On the Scannet dataset, the method achieved a Pose Error AUC (@5°) of 30.9%, significantly outperforming the best prior iterative methods (25.7%).
- Robustness: In multi-session tests on the ETH3D dataset, the proposed system succeeded across all test groups. In contrast, ORB-SLAM3 failed on 60% of the groups due to initialization failures or lost feature tracks.
- Precision: On the EuRoC-MAV drone dataset, the system demonstrated a 12x reduction in error compared to ORB-SLAM3 (0.022m vs 0.284m RMSE).
Conclusion and Future Work
By integrating classical geometric principles into a differentiable deep-learning framework, this work provides a powerful solution for multi-session SLAM. While the system currently depends on external retrieval methods (like NetVLAD) to identify candidate pairs for alignment, it sets a new benchmark for robustness in disjoint trajectory stitching. Future work may explore the integration of inertial data (IMU) to stabilize absolute scale and further optimize computational efficiency for real-time deployment on resource-constrained hardware.
Summary updated