Accelerating path tracing rendering of mesh geometry using neural models

BOGATAJ, ALJAŽ (author), Bohak, Ciril (mentor)

Neural representations have emerged as a new paradigm for representing shapes in rendering, geometric modelling and simulation. Compared to traditional representations, they can be flexibly incorporated into differentiable pipelines and be used to compress model geometry and textures. Although recent improvements to neural representations make it possible to compress model attributes while capturing its fine details, these representations are unsuitable for real-time pipelines. We propose a hybrid polygonal-neural representation suitable for efficient evaluation in path-traced algorithms. Our approach consists of a base low polygonal mesh, a rough bounding box around the final object and a set of neural networks inferring underlying model properties from ray-triangle intersections. Our architecture can be easily integrated into existing pipelines by updating the mesh intersection method to infer a simple neural network upon ray intersection with the base model. Since neural models are small and the base low poly model contains an order of magnitude fewer triangles than the original model, our method shows the potential to speed up intersection testing in ray tracing algorithms. We demonstrate the effectiveness of our approach by compressing the original model size by 4.74 times, reducing the number of triangles by 37.19 times while achieving satisfactory visual results.

• Language: English
• Keywords: Neural representation, path tracing, physically based rendering
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science ; FMF - Faculty of Mathematics and Physics
• Year: 2022
• PID: 20.500.12556/RUL-136773
• COBISS.SI-ID: 108701955
• Publication date in RUL: 20.05.2022

Secondary language

• Language: Slovenian
• Title: Pohitritev upodabljanja mrežne geometrije s sledenjem poti z uporabo nevronskih modelov

Abstract

Nevronske reprezentacije so nov način predstavitve modelov pri upodabljanju, geometrijskem modeliranju in simulaciji. V primerjavi s tradicionalnimi predstavitvami jih je mogoče fleksibilno vključiti v diferencialne cevovode in uporabiti za kompresijo geometrije in tekstur modela. Čeprav nedavne izboljšave nevronskih predstavitev omogočajo kompresijo modela in ohranjanje njegovih podrobnosti, so te predstavitve neprimerne za realno-časovne cevovode. V tem delu predstavimo hibridno poligonalno-nevronsko predstavitev, ki je primerna za učinkovito uporabo v algoritmih za sledenje poti. Naš pristop je sestavljen iz osnovne preproste poligonalne mreže, ki služi kot grob okvir okoli originalnega objekta in množice nevronskih mrež, ki kodirajo atribute osnovnega modela na podlagi podatkov o presečišču žarkov in trikotnikov. Našo arhitekturo je mogoče zlahka integrirati v obstoječe cevovode s posodobitvijo presečiščne metode. Ker so nevronski modeli majhni in je osnovni model sestavljen iz majhnega števila primitivov, naša metoda kaže potencial za pospešitev testiranja presečišč v algoritmih za sledenje žarkov. Učinkovitost našega pristopa dokazujemo s tem, da prvotno velikost modela zmanjšamo 4,74-krat, število trikotnikov zmanjšamo 37,19-krat in hkrati dosežemo zadovoljive vizualne rezultate.

• Keywords: Nevronske reprezentacije, sledenje poti, fizikalno osnovano upodabljanje