Point-Based Neural Rendering

Abstract

For decades, computer graphic designers have endeavored to reconstruct our tangible world in order to facilitate the creation of novel virtual environments. The current research in this field is developing methods that utilize neural networks to produce realistic three-dimensional reconstructions, which are generated through a computationally intensive optimization process.

While there are methods that employ, for instance, triangles, voxels, or other implicit representations to model geometry, this project is centered on the utilization of point clouds. Due to their flexibility, they are well suited to representing even complex scenes with a high level of detail. The employment of neural networks facilitates further enhancement of the quality. Consequently, this line of inquiry concentrates principally on two domains. Initially, there is an emphasis on optimizing the efficiency and quality of image generation from point clouds. Secondly, there is an emphasis on improving current network architectures. The long-term objective is to utilize the reconstructed models in interactive real-time applications. For instance, it is conceivable that smartphones could be used not only to capture and view simple images but also detailed 3D reconstructions.