3D reconstruction of underwater objects using NeRF

Abstract

This paper presents a method for reconstructing the 3D shape of underwater objects using Neural Radiance Fields. Underwater reconstruction is often hindered by image degradation caused by environmental noise. To address this, we utilize the model’s ability to generate continuous volumetric representations from sparse and incomplete data, selectively training it on low-noise underwater images to reduce distortion effects. While standard view synthesis models offer high-quality and realistic reconstructions, they require long training times, limiting their usability in real-time applications. To overcome this, we employ Instant Neural Graphics Primitives, which significantly reduces training time while maintaining visual fidelity. For validation, we captured image data of the same anchor object in both terrestrial and underwater environments and trained the models separately. Reconstruction quality was assessed using Peak Signal-to-Noise Ratio, Structural Similarity Index Measure, and Learned Perceptual Image Patch Similarity metrics. The terrestrial data achieved an average Peak Signal-to-Noise Ratio of 28.51, Structural Similarity Index Measure of 0.97, and Learned Perceptual Image Patch Similarity of 0.07, while underwater data yielded 23.23, 0.71, and 0.15, respectively. Despite the inherent challenges in underwater imaging, the results demonstrate that the proposed method can achieve natural and reliable 3D reconstructions in both settings.