HR-NeRF: advancing realism and accuracy in highlight scene representation.

Abstract

NeRF and its variants excel in novel view synthesis but struggle with scenes featuring specular highlights. To address this limitation, we introduce the Highlight Recovery Network (HRNet), a new architecture that enhances NeRF's ability to capture specular scenes. HRNet incorporates Swish activation functions, affine transformations, multilayer perceptrons (MLPs), and residual blocks, which collectively enable smooth non-linear transformations, adaptive feature scaling, and hierarchical feature extraction. The residual connections help mitigate the vanishing gradient problem, ensuring stable training. Despite the simplicity of HRNet's components, it achieves impressive results in recovering specular highlights. Additionally, a density voxel grid enhances model efficiency. Evaluations on four inward-facing benchmarks demonstrate that our approach outperforms NeRF and its variants, achieving a 3-5 dB PSNR improvement on each dataset while accurately capturing scene details. Furthermore, our method effectively preserves image details without requiring positional encoding, rendering a single scene in ~18 min on an NVIDIA RTX 3090 Ti GPU.