Trodimenzionalna rekonstrukcija čovjeka iz jedne slike

master's thesis

• Title (english): 3D Human Reconstruction from a Single Image
• Author: Matej Đanić
• Mentor: Tomislav Pribanić (mentor)
• Committee member: Tomislav Pribanić (predsjednik povjerenstva)
• Committee member: Marko Subašić (član povjerenstva)
• Committee member: Hrvoje Džapo (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2021-07-06, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

U diplomskom radu naveden je utjecaj napretka metoda dubokog učenja na zastupljenost njihove upotrebe u područjima računalnog vida i problematike 3D rekonstrukcije. Navedene su prednosti i mane upotrebe dosadašnjih rješenja za 3D rekonstrukciju koja ne koriste metode dubokog učenja. Objašnjena je važnost digitalizacije ljudskog tijela i njegove primjene u područjima virtualne i proširene stvarnosti. Opisuju se rješenja koja koriste konvolucijske neuronske mreže za problematiku 3D rekonstrukcije ljudskog tijela te njihova podjela prema obliku i načinu prikupljanja skupova podataka za treniranje. Navedene su arhitekture slojeva odabranih neuronskih mreža i prikupljenih podataka za treniranje. Također je opisan postupak predobrade 3D modela i slika u svrhu generiranja sintetiziranih podataka za treniranje mreže. Mreže, implementirane u programskom jeziku Python upotrebom biblioteke Tensorflow, eksperimentalno su evaluirane u procesu treniranja, a rezultati analizirani i uspoređeni ovisno o tipu korištenih sintetiziranih podataka.

Abstract (english)

This Master's thesis presents the impact of the progress of deep learning methods on the prevalence of their use in the fields of computer vision and the task of 3D reconstruction. The advantages and disadvantages of using existing solutions for 3D reconstruction without deep learning methods are listed. The importance of digitalization of the human body and its application in the fields of virtual and augmented realities is explained. Solutions using convolutional neural networks for the task of 3D human body reconstruction and their division according to the form and method of dataset collection is described. Layer architectures of selected neural networks and collected datasets for training are listed. Moreover, the method of preprocessing 3D models and images for the purpose of generating synthesized data for training the networks is described. Networks, implemented in the Python programming language using the Tensorflow library, are experimentally evaluated in the process of training, and the results are analyzed and compared according to the use type of synthesized data that is used.

• Keywords: 3D model čovjeka; 3D rekonstrukcija; RGB-D podatci; više pogleda; vokseli; duboko učenje; konvolucijska neuronska mreža; arhitektura slojeva; treniranje; validacija; skupovi podataka THuman i LSUN; predobrada podataka; funkcija gubitka; metoda optimizacije
• Keywords (english): 3D human model; 3D reconstruction; RGB-D data; multi-view; voxels; deep learning; convolutional neural network; layer architecture; training; validation; THuman and LSUN datasets; data preprocessing; loss function; optimization method
• Language: croatian
• URN:NBN: urn:nbn:hr:168:726200
• 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: 2023-01-13 14:51:14