Analysis of refraction aware Neural Radiance Fields for 3D reconstruction of through the water scenes

Abstract

Neural Radiance Fields (NeRFs) synthesize novel views based on images acquired from different camera positions to represent 3D scenes. However, since they assume linear light paths, they are unsuitable for underwater environments where refraction causes nonlinear ray trajectories, resulting in blurred scene reconstructions due to the absence of physical light path modeling. Developments, such as NeRFrac, try to explicitly model refractive surfaces by incorporating Snell’s law into NeRF frameworks. Nevertheless, the predominant objective within the computer vision community, the generation of high-quality renderings, assessed through metrics, e.g. including the Peak Signal-to-Noise Ratio (PSNR) persists. However, the main task in photogrammetric and geospatial applications is geometric reconstruction in the form of 3D point clouds. Therefore, this work investigates the possibilities of extracting 3D point clouds from refraction-aware NeRF implementations, specifically evaluating the NeRFrac codebase with the use of nine images.