Indoor scene generation method using radiance fields and super-resolution

• Main Author: Yang, Yida
• Other Authors: Liu Ziwei
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Indoor scene generation; Radiance fields; Super-resolution; Generative adversarial network
• Online Access: https://hdl.handle.net/10356/175309
• Institution: Nanyang Technological University

Indoor scene generation in the digital realm has garnered significant attention within the computer vision domain, offering diverse applications ranging from architectural visualization to virtual reality experiences and gaming environments. Traditional methods relying on manual 3D modeling are time-consuming and lack scalability. Recent advancements, particularly the introduction of Neural Radiance Field (NeRF), have shown promise in representing indoor scenes comprehensively and synthesizing novel views. This Final Year Project (FYP) proposes a method combining NeRF-based scene generation with single image super-resolution methods. By leveraging a generative adversarial network (GAN) with radiance field and employing convolutional neural networks (CNNs) for super-resolution, the method aims to enhance the realism and resolution of generated indoor scenes. Experimental results demonstrate improvements over baseline models, although issues regarding consistency are noted and discussed.