A study of the effectiveness and computational cost of computer vision models for outdoor segmentation

Implementing a Computer Vision system able to recognize outdoor scenarios presents inherent difficulties related to the very high variability of environmental factors that make the task challenging and even poorly explored. The use case presented in this thesis is intended to focus on these types of scenarios, even more specifically with those that can be visited by lawn mowing robots, to compare and analyze the effectiveness and computational cost of various Computer Vision models specialized in the task of instance segmentation. In order to do this, both Convolutional Neural Network (CNN) and Vision Transformer (ViT) will be tested on multiple datasets representing outdoor environments with a series of experiments designed to test the capabilities and the limitations of these models.