Overview

Code implementation of Spatial Annealing Smoothing for Efficient Few-shot Neural Rendering
Comparisons:

This code is implemented based on TriMipRF.

Installation

First, create a new sanerf environment:

conda create -n sanerf python==3.8

Next, activate the environment:

conda activate sanerf

Install the following dependency:
PyTorch (1.13.1 + CUDA 11.6)
nvdiffrast
tiny-cuda-nn

pip install -r requirements.txt

DATA

nerf_synthetic dataset

Please download the nerf_synthetic dataset from the NeRF official project.

Reproduce SANeRF's results

bash ./scripts/reproduce_sanerf.sh

The reproduced results have an error margin of 0.05dB.

Reproduce baseline's results

Replace method=sanerf with method=trimiprf in "./scripts/reproduce_sanerf.sh" , then run:

bash ./scripts/reproduce_sanerf.sh

Citation

If you find our work useful, please cite it as

@misc{xiao2024spatial,
      title={Spatial Annealing Smoothing for Efficient Few-shot Neural Rendering}, 
      author={Yuru Xiao and Xianming Liu and Deming Zhai and Kui Jiang and Junjun Jiang and Xiangyang Ji},
      year={2024},
      eprint={2406.07828},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}