Generation and Rendering of Transparent and Translucent Objects

Bachelor Thesis

Author: Michal Kubirita
Supervisor: Mgr. Lukáš Gajdošech
Consultant: doc. RNDr. Martin Madaras, PhD.

Motivation

Training of robust neural networks is based on the access to the training data. A convolutional neural network can be trained in a deep learning manner using synthetic or real captured annotated data to perform point cloud processing tasks, such as filtering the point clouds or pose estimation of an object. If there is no available real annotated training dataset, a synthetic one can be rendered and used. The main scope of the thesis would be to propose a rendering pipeline for synthetic scans composed of transparent and translucent objects. The generated data will be later used to train a neural network for point cloud processing. The translucent and transparent objects should be parametrized in order to be randomized in the scene. In an optimal scenario, the rendering of the scene should be performed in real-time, therefore Unreal Engine should be used. Alternatively, a Blender with LuxCoreRenderer can be used.

Aim

• Create models of translucent and transparent objects that can be parametrized and added to a scene using a script.
• Explore the possibilities of rendering engines, e.g. Unreal Engine, Blender’s LuxCoreRender, for their capabilities of rendering translucent and transparent objects.
• Prioritize faster, real-time rendering. Therefore, focus mainly on Unreal Engine (NVIDIA UE branch).
• Render the scene with caustics and compare the results from UE and Blender

Literature

Rui Wang, Wei Hua, Yuchi Huo, Hujun Bao 2022, Real-time Rendering and Editing of Scattering Effects for Translucent Objects, https://doi.org/10.48550/arXiv.2203.12339