Kvantifikacija stabilnosti lokalizacije objekta pravokutnim okvirom u problemu detekcije objekta

master's thesis

• Title (english): Quantification of stability of object localization with a rectangular box in the problem of object detection
• Author: Roko Torbarina
• Mentor: Luka Grubišić (mentor)
• Committee member: Luka Grubišić (predsjednik povjerenstva)
• Committee member: Goran Igaly (član povjerenstva)
• Committee member: Boris Muha (član povjerenstva)
• Committee member: Ozren Perše (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Mathematics); Zagreb
• Defense date and country: 2023-09-27, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Mathematics

Abstract

U teorijskom dijelu rada, opisane su neuronske mreže i njihove glavne komponente. Postepenom izgradnjom i uvođenjem novih koncepata, dolazi se do konvolucijskih neuronskih mreža te njihove uporabe u problemu detekcije objekta. Predstavljena je kvantifikacija stabilnosti lokalizacije objekta pravokutnim okvirom te su navedene najpopularnije metode kvantifikacije stabilnosti poput Bayesovih neuronskih mreža i Monte Carlo isključivanja. U praktičnom dijelu rada, nakon predstavljanja skupa podataka, opisana je metoda korištenja Gaussove negativne logaritamske vjerodostojnosti kao funkcije gubitka za učenje kvantifikacije stabilnosti lokalizacije objekta. Opisana je RetinaNet arhitektura korištena za treniranje modela s navedenom funkcijom greške. Provedena su različita testiranja te su dobiveni obečavajući rezultati.

Abstract (english)

In the theoretical part of the thesis, neural networks and their main components are described. By gradually building and introducing new concepts, we arrive to convolutional neural networks and their use in object detection problem. Quantification of bounding box stability is introduced and the most popular methods of stability quantification, such as Bayesian neural networks and Monte Carlo dropout are discussed. In the practical part of the paper, after the presentation of the data set, the method of using Gaussian negative log-likelihood as a loss function for learning the quantification of object localization stability is presented. The RetinaNet architecture used to train the model with the specified loss function is described. Various tests were carried out and promising results were obtained.

• Keywords: neuronske mreže; konvolucijske neuronske mreže; Bayesove neuronske mreže; Monte Carlo isključivanja
• Keywords (english): neural networks; Bayesian neural networks; Monte Carlo dropout
• Language: croatian
• URN:NBN: urn:nbn:hr:217:231914
• Study programme: Title: Computer Science and Mathematics; Study programme type: university; Study level: graduate; Academic / professional title: sveučilišni magistar računarstva i matematike (sveučilišni magistar računarstva i matematike)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2024-02-06 13:12:00