Check out our paper Specularity in NeRFs: A Comparative Study of Ref-NeRF and NRFF (2025), published in the MLBriefs 2024 workshop of the Image Processing On Line journal (IPOL).
This demo compares two models for specular reflections: Ref-NeRF [1] and NRFF [2]. The comparison is performed across multiple views in RGB, normal maps, and three evaluation metrics (PSNR, SSIM, and LPIPS).
You can try the demo here: DEMO
For this evaluation, two objects were selected from the Shiny Blender dataset, and an additional object was specifically designed, which can be found here.
Figure 1: Sample comparisons for anisotropic object. The first row shows GT, NRFF, and Ref-NeRF RGB. The second row shows NRFF and Ref-NeRF Normals.
To run the demo on your local machine:
-
Clone the repository:
git clone https://github.com/AlbertBarreiro/Albert_demo_mlbriefs4.git cd Albert_demo_mlbriefs4 -
Ensure the necessary dependencies are installed:
pip install -r requirements.txt
-
Set the bin path:
bin="$(pwd)" -
Run the demo with the desired dataset file and azimuth angle (must be between -160 and 160):
bash main.sh <dataset> (e.g., dataset_toaster, dataset_ani, dataset_ball) <azimuth_angle>
Example:
bash main.sh dataset_toaster 30
If you find this useful, please cite:
@article{barreiro2025specularity,
title={Specularity in NeRFs: A Comparative Study of Ref-NeRF and NRFF},
author={Barreiro, Albert and Mar{\'\i}, Roger and Redondo, Rafael and Haro, Gloria and Bosch, Carles and Berga, David},
journal={Image Processing On Line},
year={2025}
}
This code is released under the MIT License. See LICENSE for details.
[1] D. Verbin, P. Hedman, B. Mildenhall, T. Zickler, J. T. Barron, and P. P. Srinivasan, “Ref-nerf: Structured view-dependent appearance for neural radiance fields,” CVPR, 2022. Ref-NeRF Info Page
[2] K. Han and W. Xiang, “Multiscale tensor decomposition and rendering equation encoding for view synthesis,” CVPR, 2023. NRFF Info Page