Hyperparameter Optimization for SLAM: An Approach For Enhancing ORB-SLAM2's Performance

Abstract

Simultaneous Location and Mapping (SLAM) has been a well-pursued research area for computer vision and robotics. Robustness and performance are fields that address the efficiency of SLAM solutions. Hyperparameter Optimization (HPO) promises to find a hyperparameter set that displays the lowest error within a validation set. This thesis aims to devise a methodology that applies HPO to SLAM to reduce the absolute trajectory error produced and to increase performance by building a more accurate map. Specifically, it investigates whether the proposed methodology impacts error reduction on ORB-SLAM2. We train model-free, population-based algorithms in a modified KITTI benchmark to obtain an initial set of possible configurations and test them against model-free, search-based baseline algorithms. We used a combination of 20 modified and unaltered sequences for performance evaluation. Four evaluation metrics (optimality, proximity, under-performance, and success rates) determine the efficacy of each candidate configuration. The proposed methodology outperformed a default configuration execution with an 80% success rate. The results promise case-specific executions. However, we could not find a universal hyperparameter set to reduce error in all test cases. The proposed methodology has a simple implementation, is cost-effective, does not need an expert tuner, and shows up to 60% error reduction.