Adaptacija domene za semantičku segmentaciju slika

master's thesis

• Title (english): Domain adaptation for semantic segmentation
• Author: Leon Bertol
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Marko Čupić (član povjerenstva)
• Committee member: Leonardo Jelenković (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2023-07-14, Croatia

Abstract

U ovom radu demonstrirali smo dvije različite metode za adaptaciju domene za semantičku segmentaciju prometnih scena. Kao izvorišne, odnosno sintetičke podatke za treniranje, koristili smo skup podataka dobivenih iz GTA5 i CARLA smiulatora. Ciljana domena, odnosno ona nad kojom smo testirali modele, čini skup podataka Cityscapes. Dok prva metoda suparničkim učenjem nastoji ciljanu domenu prilagoditi izvorišnoj, druga metoda prilagođava sam model kako bi dobro generalizirao na ciljanoj domeni koristeći tehnike samostalnog učenja. Kako bismo ispitali i bolje razumjeli što sve ima utjecaj na adaptaciju domene, isprobali smo različite arhitekture, hiperparametre te dopunske metode i usporedili dobivene rezultate.

Abstract (english)

In this paper, we demonstrated two different methods for domain adaptation for semantic segmentation of traffic scenes. As source, or synthetic data for training, we used a set of data obtained from GTA5 and CARLA simulators. The target domain, that is, the one on which we tested the models, is the dataset Cityscapes. While the first method tries to adapt the target domain to the original by adversarial learning, the second method adapts the model itself to generalize well on the target domain using self-learning techniques. In order to examine and better understand what influences domain adaptation, we tried different architectures, hyperparameters and supplementary methods and compared the obtained results.

• Keywords: Adaptacija domene; semanticka segmentacija; transformer; CycleGAN; duboko učenje
• Keywords (english): Domain adaptation; semantic segmentation; transformer; CycleGAN; deep learning
• Language: croatian
• URN:NBN: urn:nbn:hr:168:229972
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-10-11 08:26:06