Comparison of software implementations of the Bundle Adjustment algorithm for the SLAM problem

This paper investigates the problem of Simultaneous Localization and Mapping (SLAM) with a focus on the optimization stage of Bundle Adjustment (BA). This stage is crucial for accurately reconstructing the 3D structure of a scene and estimating camera parameters, which directly impact subsequent computer vision tasks such as tracking, depth map generation, and reconstruction. Modern software libraries implementing the BA algorithm are reviewed, including Ceres Solver, GTSAM, and SymForce. An experimental comparison of these libraries is conducted based on performance criteria such as execution speed, number of iterations, Absolute Pose Error (APE), and Cost Change. The BAL (Bundle Adjustment in the Large) dataset and three specific datasets—Ladybug, Trafalgar Square, and Dubrovnik—were used for testing. A software solution was developed to evaluate the performance of these libraries, incorporating steps such as normalization, adding random noise, and executing optimization algorithms. The visualization of solutions was performed using the Evo library. Experimental results show that SymForce provides the best balance between speed, accuracy, and convergence. It achieved the shortest execution time for the Ladybug and Trafalgar Square datasets and exhibited the lowest Absolute Pose Error across all cases. Ceres Solver proved competitive, particularly for the Dubrovnik dataset, where it outperformed SymForce in terms of speed. In contrast, GTSAM demonstrated the worst performance across all criteria, exhibiting the highest absolute errors and the longest computation times. Based on these findings, SymForce is the most suitable tool for SLAM and nonlinear optimization tasks where both speed and accuracy are critical. Ceres Solver can be effective in specific scenarios requiring maximum computational speed. Meanwhile, GTSAM lags behind its competitors, limiting its applicability in high-performance systems. These conclusions will aid in selecting the appropriate optimization libraries for practical applications in robotics, autonomous vehicles, and 3D mapping.