Comparison of the Bayesian Optimization-based Tree-structured Parzen Estimator Algorithm

This work contributes to a measurement automation project by improving its vision system, whose task is to create high-quality images and process them to aid robot navigation. One of the image processing challenges for the vision system is semantic segmentation, which is carried out by fully convolutional neural networks. The focus of this work is the hyperparameter fine-tuning of these fully convolutional neural networks. First, the basic concept of the Bayesian optimization-based Treestructured Parzen Estimator Algorithm is presented. Next, the main characteristics of the bandit-based Successive Halving algorithm are highlighted. The goal of this work is to compare the performance of these methods under the same runtime conditions. The comparison is made using internationally accepted metrics: Intersection over Union, Mean Class Accuracy, and Pixel Accuracy.