Robots have become better at understanding their surroundings and making smart decisions thanks to semantic maps. In recent years, neural networks have greatly improved their ability to perceive the world. However, there's a problem: these networks tend to be overly confident about their predictions and are not able to express the uncertainty of the model, resulting in wrong information.
To address this issue, we propose a new method in [1] to effectively characterize the fusion process into a map of semantic information coming from a semantic segmentation neural network. It uses a Bayesian approach to use the confidences of the network about the data (aleatoric uncertainty) but includes a regularization to account for overconfidence. Additionally, we leverage Bayesian neural networks to obtain the model uncertainty (epistemic uncertainty) and include this in the fusion process as Dirichlet concentration parameters. This method is able to reduce the influence of outliers and uncertain measurements to improve the quality of the final semantic map.
References
[1] D. Morilla-Cabello, L. Mur-Labadia, R. Martinez-Cantin, E. Montijano, "Robust Fusion for Bayesian Semantic Mapping," in IEEE/RSJ International Conference on Intelligent Robots and Systems, Oct. 2023. Available online at: https://arxiv.org/abs/2303.07836