Reconstruction of Nepali Statues: A Comparative Evaluation of Neural Radiance Fields and 3D Gaussian Splatting

Abstract

Digital preservation of cultural-heritage artifacts
such as statues has traditionally depended on specialized 3D
scanning hardware – Light Detection and Ranging (LiDAR)
scanners, structured-light sensors, or depth cameras – that are
costly and largely inaccessible to smaller institutions. This paper
presented an accessible reconstruction pipeline for generating
3D models of statues from ordinary multi-view 2D photographs,
combining Structure-from-Motion (SfM) camera-pose estimation
via COLMAP with two neural rendering paradigms: Nerfacto,
an accelerated Neural Radiance Field (NeRF) variant, and 3D
Gaussian Splatting (3DGS). A controlled, statue-specific comparative
evaluation of Nerfacto and 3DGS was conducted on two
independently captured, self-collected statue datasets (146 images
captured per statue, 145 retained after quality filtering), trained
under a matched 15,000-iteration budget on a single Tesla T4
GPU. Across both statues, 3DGS achieved higher reconstruction
fidelity than Nerfacto – average Peak Signal-to-Noise Ratio
(PSNR) of 20.80 dB and 11.15 dB versus Nerfacto’s 10.91 dB and
9.64 dB, with corresponding gains in Structural Similarity Index
Measure (SSIM) and Learned Perceptual Image Patch Similarity
(LPIPS) – while training 5×–7× faster (17–24 minutes versus 120
minutes per statue). Given the small sample, these results were
treated as preliminary evidence favoring 3DGS under resourceconstrained
training budgets rather than a general claim. To
the authors’ knowledge, this was among the first reported
applications of this comparison to Nepali cultural statues. A
COLMAP camera-model incompatibility between the distortionaware
OPENCV model and the distortion-free 3DGS rasterizer,
and a non-uniform default train/test split identified for one
dataset, were also documented and discussed as methodological
limitations.