3D Reconstruction based on NeRF and SDF methods: A comparative evaluation using RGB-D data

Abstract

This thesis explores the process of 3D scene reconstruction using a depth (RGB-D) camera, combined with advanced methodologies in artificial intelligence and visual computing. The research involves capturing real-world scenes using the RGB-D camera, followed by exporting each frame through multiway registration using the Open3D library to ensure accurate alignment and reconstruction. The core of this work is conducted within SDFStudio, an extension built on the NeRF studio framework, which facilitates the development and experimentation of methods involving Signed Distance Fields (SDFs). SDFs are crucial for representing 3D shapes and surfaces with precision, making them ideal for applications requiring accurate geometric computations. Leveraging the modular design and features of SDFStudio, the research implements and compares three state-of-the-art SDF-based algorithms: Neural Unsigned Distance Fields - facto (NeuS-facto), UNISURF, and MonoSDF. These methods are tested on datasets comprising depth and RGB images along with known camera parameters (poses, and intrinsic). The performance and accuracy of the algorithms are systematically evaluated by adjusting key parameters, such as SDF grid resolution, number of iterations, and learning rates, to assess their impact on 3D reconstructions quality.