X-SLAM: Scalable Dense SLAM for Task-aware Optimization using CSFD

TL;DR

X-SLAM introduces a real-time dense differentiable SLAM system that uses complex-step finite difference (CSFD) to compute derivatives without a large computation graph, enabling efficient task-aware optimization. The authors present two end-to-end frameworks, X-KF and X-EF, for downstream tasks such as outdoor camera relocalization and indoor active scanning, demonstrating improvements in accuracy and navigation efficiency on public benchmarks and real scenes. A thorough evaluation compares CSFD against forward differences and automatic differentiation, highlighting superior memory efficiency and the feasibility of high-order optimization (e.g., Hessians) in real-time settings. This work advances practical differentiable SLAM by providing a scalable, memory-conscious approach that supports high-order differentiation for improved downstream task performance.

Abstract

We present X-SLAM, a real-time dense differentiable SLAM system that leverages the complex-step finite difference (CSFD) method for efficient calculation of numerical derivatives, bypassing the need for a large-scale computational graph. The key to our approach is treating the SLAM process as a differentiable function, enabling the calculation of the derivatives of important SLAM parameters through Taylor series expansion within the complex domain. Our system allows for the real-time calculation of not just the gradient, but also higher-order differentiation. This facilitates the use of high-order optimizers to achieve better accuracy and faster convergence. Building on X-SLAM, we implemented end-to-end optimization frameworks for two important tasks: camera relocalization in wide outdoor scenes and active robotic scanning in complex indoor environments. Comprehensive evaluations on public benchmarks and intricate real scenes underscore the improvements in the accuracy of camera relocalization and the efficiency of robotic navigation achieved through our task-aware optimization. The code and data are available at https://gapszju.github.io/X-SLAM.

Figures (9)

• Figure 1: An illustration of the one-to-one association (left) and our one-to-multiple association (right). The associated global model points (gray dots) of camera ray (green dashed line) is colored in orange.
• Figure 2: The optimization process of ICP based on different optimization methods. Translation error (unit: $m$), rotation error (unit: $^\circ$) and point-to-plane ICP loss are reported.
• Figure 3: Absolute distance error (unit: $m$) during tracking. Left: frame step = 1, Mid: frame step = 2, Right: frame step = 3. High-order ICP can achieve better tracking accuracy when the camera moves faster.
• Figure 4: Camera relocalization trajectory in 7-Scenes dataset for HLoc. For all scenes, our optimization method based on X-KF can significantly enhance the relocalization accuracy. The results also demonstrate the importance of higher-order optimization, which is difficult for AD method.
• Figure 5: Optimization curves on the chess scene in 7-Scenes dataset. Three frames (No. 31, No. 61, No. 91) are optimized 50 iterations in Gradient Descent and Newton's method. Not only does Newton's method achieve a faster convergence speed, but it also prevents falling into local minima.