Advisors: Matteo Matteucci, Matteo Bellusci

Nowadays, Advanced Driver Assistance Systems (ADAS) are a common feature in new vehicles. Such critical systems may require a large quantity of data to be optimally trained, whose acquisition can be an expensive task both in terms of time and cost. This process can be simplified by generating synthetic data through a simulator, but these data might lead the ADAS components to yield results different from the ones that would have been obtained by training the system on data from the real world; discrepancy is known as sim-to-real gap. In this work, a photorealistic driving simulator has been used to generate a synthetic dataset for solving semantic image segmentation tasks, to measure the sim-toreal gap related to models trained with it. To do so, the aforementioned dataset has been used to train a deep neural network model, its performances have then been compared with the ones obtained by training the same model with Cityscapes and CamVid, two datasets acquired in the real world. These results show the existence of a performance gap between models trained on simulated data and the same models trained on real-world data, which is mainly revealed by the difficulties that the networks of the first type have in classifying the image pixels associated with the less frequent dataset labels. On the other hand, the outputs from the models trained on the synthetic dataset show that these networks can identify the general structure of the image received in input and correctly classify pixels related to the most common labels, meaning that more thorough works on datasets created with this driving simulator might produce better results.

ADAS, sim-to-real gap, semantic segmentation, vision sensing and perception, deep learning, intelligent vehicles

This folder contains the Jupyter notebooks created for supporting the thesis work:

• midgard_segmentation3.ipynb: Jupyter notebook used for the datasets experiments
• Utilities.ipynb: Jupyter notebook used to generate black and white segmentation masks, resize datasets and generate bar graphs

This folder contains the thesis document, together with the thesis' presentation