Neural radiance fields : extending supervision and spatial encoding
Loading...
Date
2025-10-13
Authors
Σίμωση, Μαρία
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Three-dimensional (3D) reconstruction is a fundamental problem in computer vision with wide applications. Recent advances have been driven by neural implicit representations, where scenes are encoded as continuous functions parameterized by neural networks. A landmark in this domain is Neural Radiance Fields (NeRF), which achieves high-fidelity novel view synthesis but suffers from high computational cost.
This thesis presents a PyTorch-based implementation of NeRF, covering the entire pipeline from dataset preprocessing and spatial encoding to network training, volume rendering, and real-time interactive visualization. Beyond a baseline reproduction of NeRF, we introduce two contributions: (i) explicit supervision of opacity in the loss function, and (ii) a novel hash encoding scheme based on the Hilbert curve.
We evaluate our implementation on the widely used nerf-synthetic benchmark, reporting results with PSNR, SSIM, and LPIPS metrics. Additionally, we explore the challenges of extending NeRF to outdoor scenes captured from video, identifying limitations that suggest avenues for future research.
Overall, this work offers a clear and practical NeRF implementation alongside experimental insights that point toward promising avenues for future development.
Description
Keywords
Neural radiance fields, NeRF, 3D reconstruction, Computer vision