A Probabilistic Framework Applied to Multispectral Image Segmentation

This paper describes a study on exploring probability densities applied to the logit values of deep models to obtain predictions with probabilistic interpretation. Kernel density estimation (KDE) is used to providing non-parametric modeling of the underlying logit-values distribution. While preliminary results show promising potential, this work lays the foundation for future research, where the combination of KDE with other advanced techniques could further enhance segmentation performance of deep models. This study reports segmentation results on a dataset collected by a UAV equipped with a multispectral camera flying over vineyards.