Understanding the Effect of Deep Ensembles in LiDAR-Based Place Recognition

Place recognition, the task of recognizing a previously visited location, has a decisive role in the autonomous driving field since it enables rough global localization in GNSS-denied environments. In the last few years, LiDAR-based place recognition and deep learning approaches achieved outstanding results also within challenging scenarios. However, the use of DNN-based methods is still limited due to the safety-critical nature of the task and the difficulty in detecting potential model failures. Determining the uncertainty of DNN-based outputs is a useful technique to discover unreliable predictions. Among the existing approaches, Deep Ensemble represents a popular sampling method to estimate epistemic uncertainty by exploiting multiple models. However, an in-depth investigation of its application for LiDAR-based place recognition is missing and only one approach has been recently proposed [22]. Our ultimate goal is to gain a deeper understanding of the strengths and weaknesses of Deep Ensemble methods. To achieve this, we propose a Deep Ensemble strategy that uses a knowledge-distillation approach and we compare it to [22] by evaluating its recall and failure detection capabilities.