Epistemic uncertainty estimation with evidential learning on semantic segmentation of underwater images

摘要

As the evolution of robotics occurs, it has become increasingly common for robotic systems to be integrated with intelligent systems. Despite the relevant advance in the area of neural networks, what happens inside a network is still not completely comprehended. The uncertainties in the neural network prediction can be epistemic, which is related to the model, or aleatory, which is related to the data. The reliability of network predictions can be increased with uncertainty estimation. Making the fusion of sensor data with trustful predictions, robotic systems could operate in more critical environments, where high accuracy is required, as well as maintaining navigation in the case of sensor fault. Underwater environments suffer from some disturbances that affect the acquisition of images due to light interference or wave movement, and these perturbations make difficult the use of neural networks for underwater robots, due to doubtful predictions. Taking this into account, this paper proposes a methodology to estimate the epistemic uncertainty with the case study in semantic segmentation for underwater images, where the behavior of uncertainty during training is observed, resulting in a qualitative analysis of epistemic uncertainty for Semantic Segmentation. The results suggest estimated epistemic uncertainty is directly connected with the parameters, amounts of epochs, size, and quality of the dataset.