Določanje osvetlitve scene z ocenjevanjem svetlobne karte

Kolar, Klemen (author), Čehovin Zajc, Luka (mentor)

Problem s katerim se ukvarjamo v diplomskem delu je določanje osvetlitve scene z napovedovanjem svetlobnega vira za uporabo v obogateni resničnosti. Predlagamo novo tehniko določanja svetlobnih kart z uporabo globokih nevronskih mrež in novo sintetično podatkovno množico za učenje modela. Svetlobne karte predstavljajo kodiranje dveh kotov potrebnih za določanje vira svetlobe v matriko vseh možnih parov le-teh. Preizkusimo in primerjamo različne arhitekture hrbtenice nevronske mreže in različne tehnike augmentacije učnih podatkov modela. Za testiranje uspešnosti model primerjamo z že preizkušenima metodama napovedovanja radianov in ločenih kotov na realni nevideni množici podatkov. Končni model je dosegel bolj točne napovedi, kot prej omenjeni tehniki.

• Language: Slovenian
• Keywords: strojno učenje, globoko učenje, obogatena resničnost, določanje osvetlitve
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2024
• PID: 20.500.12556/RUL-161576
• COBISS.SI-ID: 213358083
• Publication date in RUL: 12.09.2024

KOLAR, Klemen, 2024, Določanje osvetlitve scene z ocenjevanjem svetlobne karte [online]. Bachelor’s thesis. [Accessed 11 April 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=161576

Secondary language

• Language: English
• Title: Determining scene illumination by light map estimation

Abstract

The problem of this thesis is determining the scene illumination by predicting the light source for use in augmented reality. We propose a new technique for determining light maps using deep neural networks and a new synthetic dataset for model learning. Light maps represent the encoding of the two angles needed to determine the light source into a matrix of all possible pairs of them. We test and compare different neural network backbone architectures and different model learning data augmentation techniques. To test the performance, we compare the model with two previously tested methods for predicting radians and disjointed angles on a real unseen dataset. The final model achieved more accurate predictions than the previously mentioned techniques.

• Keywords: machine learning, deep learning, augmented reality, light estimation