Analiza generalizacije semantične segmentacije z globokimi zbirkami filtrov

Prelevikj, Marko (author), Kristan, Matej (mentor), Hlavač, Vaclav (comentor)

Mobilni robotski sistemi, ki so sposobni avtonomne navigacije v nestrukturiranih okoljih, so odvisni od njihovih modulih vida, da bi lahko bili sposobni se navigirati čez okolje. Moduli vida priskrbijo percepcijo okolice, in pogosto morajo identificirati določene predmete, ki nas zanimajo. Identifikacija nastane tako da določene segmente slik klasificira v enem izmed vnaprej naučenih razredov. Na področju računalniškega vida obstaja veliko postopkov semantične segmentacije, ki poročajo izjemne rezultate. Vendar so ti postopki naučeni samo na določenih podatkovnih zbirkah, ki niso nujno medsebojno odvisni z različnimi prizorišči, ki jih mobilni robot opazi. Da bi preverili sposobnost podatkovne zbirke prenesti svoje znanje na novi domeni bomo preiskovali kvaliteto generalizacije njenih razredov. Preverili bomo prenos znanja specifičnega postopka semantične segmentacije, ki smo ga prilagodili našim potrebam.

• Language: English
• Keywords: semantična segmentacija, konvolucijske nevronske mreže, zaznavanje tekstur, prenos znanja
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2017
• PID: 20.500.12556/RUL-92730
• Publication date in RUL: 30.06.2017

PRELEVIKJ, Marko, 2017, Analiza generalizacije semantične segmentacije z globokimi zbirkami filtrov [online]. Bachelor’s thesis. [Accessed 2 April 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=92730

Secondary language

• Language: Slovenian
• Title: Generalization analysis of semantic segmentation with deep filter banks

Abstract

Mobile robotic systems capable of autonomous navigation in non-structured environments depend on their vision module in order to safely navigate through the environment. The vision module provides perception of the surrounding area and it is often required to identify particular objects of interest, which is done by classifying image segments into pre-learned semantic classes. There are many methods which provide remarkable semantic segmentation results, but unfortunately only on specific datasets, which are not necessarily correlated to the scenes observed by a mobile robot. To verify the dataset's capability of transferring knowledge to a new domain we explore how well it generalises its classes. We examine the transfer of knowledge on a specific semantic segmentation method, which we adjust to best fit our needs.

• Keywords: semantic segmentation, transfer of knowledge, convolutional neural networks, texture recognition