hcp6897 / neural-singular-hessian

Code of Neural-Singular-Hessian: Implicit Neural Representation of Unoriented Point Clouds by Enforcing Singular Hessian. ACM Transactions on Graphics (SIGGRAPH Asia 2023).

Neural-Singular-Hessian

Code of Neural-Singular-Hessian: Implicit Neural Representation of Unoriented Point Clouds by Enforcing Singular Hessian. ACM Transactions on Graphics (SIGGRAPH Asia 2023).

Project Page  / Arxiv  / Video

1. Requirements

Our cod uses PyTorch.

The code was tested with Python 3.8, torch 1.31.1, CUDA 11.6 on Ubuntu 18.04.

Using conda to create the environment and activate it.

conda env create -f env.yaml
conda activate neural_singular

2. Overfitting Single Shape

1. Put your data to ./data/sdf/input, some data already exists

2. Switch to the folder surface_reconstruction, run ./run_sdf_recon.py to reconstruct the surface. The script will reconstruct all shapes under the ./data/sdf/input (*.xyz and *.ply) files

cd surface_reconstruction
python run_sdf_recon.py

Results mesh are saved in ./surface_reconstruction/log/sdf/result_meshes

3. Shape Space Learning with DFaust

We use pytorch-lightning for shape space shape learning. Noting that the newest version may not be compatible with our code.

3.1 Data Preparation

Refer to DOGNN to prepare your data.

Change 'dataset_path', 'train_split_path' and 'test_split_path' in shapespace/shapespace_dfaust_args.py to your DOGNN path based default settings

3.2 Train

You may need to tune the number of devices in pl_conv_train_shapespace.py based on the doc of pytorch-lightning

cd shapespace
python pl_conv_train_shapespace.py

3.3 Test and Fine-tuning

The code first loaded the model in the logdir, and then inference directly. After that, the code fine-tuning the network and outputs the final result

cd shapespace
python pl_conv_finetune_shapespace.py

3.4 Pre-trained models

We use 8 RTX 3090 for training. The pre-trained models already exist under the shapespace/log_conv_all_half_eikonal_1e-4_amsgrad_200_epoch_cos_1e-6_1500_2. After preparing data, follow 3.3 to run the code

4. Evaluation

For evaluation of results, POCO has provided a great script based on ConvONet. Please refer to their code.

Acknowledgements

This code is heavily based of DiGS and idf.

Thanks to their impressive work.

Bibtex

@article{zixiong23neuralsingular,
author = {Zixiong Wang, Yunxiao Zhang, Rui Xu, Fan Zhang, Pengshuai Wang, Shuangmin Chen, Shiqing Xin, Wenping Wang, Changhe Tu},
title = {Neural-Singular-Hessian: Implicit Neural Representation of Unoriented Point Clouds by Enforcing Singular Hessian},
year = {2023},
journal = {ACM Transactions on Graphics (TOG)},
volume = {42},
number = {6},
doi = {10.1145/3618311},
publisher = {ACM}
}