Impact of Loss Functions on the Training of LiDAR-based Place Recognition Models

Abstract

Place Recognition is a fundamental task in mobile robotics and autonomous systems, enabling vehicles to navigate and perform tasks in previously visited environments. LiDAR-based Place Recognition has become increasingly popular due to its robustness to changes in illumination, weather conditions, and dynamic objects. One critical aspect of training effective Place Recognition models is the selection of an appropriate loss function. In this paper, we investigate the impact of various loss functions on the training and evaluation of LiDAR voxel-based Place Recognition. To the best of our knowledge, no previous works have compared the performance of different loss functions in this context. We compare the performance of trained models on popular public datasets Oxford RobotCar and NCLT. We also tested them on the data collected in a real-world scenario. The results prove that selecting an effective loss function results in improving the performance of the trained model.