Unsigned Orthogonal Distance Fields (UODFs)

This repository is the official implementation of our CVPR 2024 paper:

Unsigned Orthogonal Distance Fields: An Accurate Neural Implicit Representation for Diverse 3D Shapes.

Links: [Project Page] [Paper] [Video(bilibili)] [Video(youtube)]

In addition, this repository provides the implementation of the newly updated work:

Unsigned Orthogonal Distance Fields: A Neural Implicit Representation for Fast and Accurate Reconstruction of Diverse 3D Shapes.

FastUODFs' Links: [Project Page (Coming soon)] [Video(bilibili)] [Video(youtube)]

In this journal manuscript, we significantly improve the speed of surface point estimation by introducing single-point prediction and tri-plane encoding.

Core idea in one sentence. Different from SDF and UDF, each point in UODFs defines the distances along the three orthogonal directions (LR, FB, UD in the figure) and directly infers the nearest surface points (simulating the characteristics of the laser), avoiding the introduction of interpolation errors and improving the reconstruction accuracy.

Visualization

Dragon	Sliced fish in box	T-shirt

PS: All reconstruction results are free of fine-tuning.

Installation: Setup the environment

The version of PyTorch we use is 1.12.0 and the operating system is Ubuntu 20.04.

Create environment

conda create --name uodf python=3.8
conda activate uodf
pip install --upgrade pip

Dependencies

Insatll Pytorch with CUDA(this repo has been tested with CUDA 11.3). For CUDA 11.3:

pip install torch==1.12.0+cu113 torchvision==0.13.0+cu113 torchaudio==0.12.0 --extra-index-url https://download.pytorch.org/whl/cu113

Setup

Python 3 dependencies:

• pyngpmesh
• trimesh
• h5py
• pymeshlab
• scipy
• tqdm
• matplotlib
• pandas
• tensorboard

To install the dependencies, begin by navigating to the third_party directory and proceed to install pyngpmesh, which is a library separated from instant-ngp.

cd third_party/ngpmesh
pip install .
cd ../../

then

pip install -r requirements.txt

DataSet

You can use the example dataset in datasets/thingi32_normalization/ or you can put your custom datasets in datasets/ directory.

Configrations

All configurable settings are accessible within the scripts/utils/options.py

Run

To train a single model, you can use it.

python scripts/train_main.py

or if you need to test, you can use the following command.

python scripts/test_main.py

To generate fused point clouds and mesh, you can utilze the following:

python scripts/PostProcessing.py

If you want to use the single-point(S.P.) based GEP estimation method, you can use the following command.

python scripts/FastUODF/test_single.py

After that, you can utilize the following command after setting args.use_pcd_post_processing to False at line 75:

python scripts/PostProcessing.py

Evaluation

python scripts/eval_cd.py

Citation

@InProceedings{UODFs,
  title={Unsigned Orthogonal Distance Fields: An Accurate Neural Implicit Representation for Diverse 3D Shapes},
  author={Lu, Yujie and Wan, Long and Ding, Nayu and Wang, Yulong and Shen, Shuhan and Cai, Shen and Gao, Lin},
  booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2024},
  page={20551-20560}
}