This repository is the official implementation of our paper in IEEE International Conference on Multimedia and Expo (ICME 2025, oral presentation):
Neural Implicit Reconstruction and Fast Rendering Based on Dual Spherical Shell
Authors: Zijian Wang, Yuqi Liu, Yan Zhao, Binghao Wang, Shen Cai*, Yanting Zhang.
Links: [Project Page] [Video(bilibili)] [Video(Youtube)]
Given a number of pre-computed concentric spheres, local SDF fitting within DSS is enabled, and early termination as well as parallel sphere tracing are facilitated for more efficient SDF rendering.
- High-fidelity Reconstruction: low reconstruction error, in terms of chamfer distance.
- Memory and Storage Efficiency: small number of pre-computed geometric primitives.
- Rendering acceleration: early termination and parallel sphere tracing (brings about 40% speed-up).
- NeuS improvement: new sampling strategy; improvements in both accuracy and speed.
We use Thingi10k and NeRF synthetic datasets, both of which are available from their official websites.
conda env create -f environment.yml
conda activate kaolin_test
pip install torch==1.8.0+cu111 torch-cluster==1.5.9 torch-geometric==1.4.1 torch-scatter==2.0.6 torch-sparse==0.6.10 torch-spline-conv==1.2.1
cd ./submodules/miniball
python setup.py install
cd ..
cd ./kaolin_sphere-0.9.1
python setup.py develop
cd ..
cd ./libigl/python
python setup.py
cd ..
cd ..
cd ./geolab-copy
cmake . -B build
cmake --build build
python train.py
python eval.py
python eval_ssim.py
This code includes code derived from 3 third-party libraries:
@inproceedings{Wang2025DSS,
title={Neural Implicit Reconstruction and Fast Rendering Based on Dual Spherical Shell},
author={Wang, Zijian and Liu, Yuqi and Zhao, Yan and Wang, Binghao and Cai, Shen and Zhang, Yanting},
booktitle={IEEE International Conference on Multimedia and Expo (ICME)},
year={2025},
}