Neural Geometric Level of Detail: Real-time Rendering with Implicit 3D Surfaces

Official code release for NGLOD. For technical details, please refer to:

Neural Geometric Level of Detail: Real-time Rendering with Implicit 3D Surfaces
Towaki Takikawa*, Joey Litalien*, Kangxue Xin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler
In Computer Vision and Pattern Recognition (CVPR), 2021 (Oral)
[Paper] [Bibtex] [Project Page]

If you find this code useful, please consider citing:

@article{takikawa2021nglod,
    title = {Neural Geometric Level of Detail: Real-time Rendering with Implicit {3D} Shapes}, 
    author = {Towaki Takikawa and
              Joey Litalien and 
              Kangxue Yin and 
              Karsten Kreis and 
              Charles Loop and 
              Derek Nowrouzezahrai and 
              Alec Jacobson and 
              Morgan McGuire and 
              Sanja Fidler},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year = {2021},
}

Directory Structure

sol-renderer contains our real-time rnedering code.

sdf-net contains our training code.

Within sdf-net:

sdf-net/lib contains all of our core codebase.

sdf-net/app contains standalone applications that users can run.

Getting started

Python dependencies

The easiest way to get started is to create a virtual Python 3.8 environment:

conda create -n nglod python=3.8
conda activate nglod
pip install --upgrade pip
pip install -r ./infra/requirements.txt

The code also relies on OpenEXR, which requires a system library:

sudo apt install libopenexr-dev 
pip install pyexr

To see the full list of dependencies, see the requirements.

Building CUDA extensions

To build the corresponding CUDA kernels, run:

cd sdf-net/lib/extensions
chmod +x build_ext.sh && ./build_ext.sh

The above instructions were tested on Ubuntu 18.04/20.04 with CUDA 10.2/11.1.

Note. If you wish to use CUDA 10.X, you must install the appropriate CuPy module (e.g. pip install cupy-cuda102. Default: 11.1).

Training & Rendering

Note. All following commands should be ran within the sdf-net directory.

Download sample data

To download a cool armadillo:

wget https://raw.githubusercontent.com/alecjacobson/common-3d-test-models/master/data/armadillo.obj -P data/

To download a cool matcap file:

wget https://raw.githubusercontent.com/nidorx/matcaps/master/1024/6E8C48_B8CDA7_344018_A8BC94.png -O data/matcap/green.png

Training from scratch

python app/main.py \
    --net OctreeSDF \
    --num-lods 5 \
    --dataset-path data/armadillo_normalized.obj \
    --raw-obj-path data/armadillo.obj \
    --epoch 250 \
    --exp-name armadillo

This will populate _results with TensorBoard logs.

Rendering the trained model

If you set custom network parameters in training, you need to also reflect them for the renderer.

For example, if you set --feature-dim 16 above, you need to set it here too.

python app/sdf_renderer.py \
    --net OctreeSDF \
    --num-lods 5 \
    --pretrained _results/models/armadillo.pth \
    --render-res 1280 720 \
    --shading-mode matcap \
    --lod 4

By default, this will populate _results with the rendered image.

If you want to export a .npz model which can be loaded into the C++ real-time renderer, add the argument --export path/file.npz. Note that the renderer only supports the base Neural LOD configuration (the default parameters with OctreeSDF).

Core Library Development Guide

To add new functionality, you will likely want to make edits to the files in lib.

We try our best to keep our code modular, such that key components such as trainer.py and renderer.py need not be modified very frequently to add new functionalities.

To add a new network architecture for an example, you can simply add a new Python file in lib/models that inherits from a base class of choice. You will probably only need to implement the sdf method which implements the forward pass, but you have the option to override other methods as needed if more custom operations are needed.

By default, the loss function used are defined in a CLI argument, which the code will automatically parse and iterate through each loss function. The network architecture class is similarly defined in the CLI argument; simply use the exact class name, and don't forget to add a line in __init__.py to resolve the namespace.

App Development Guide

To make apps that use the core library, add the sdf-net directory into the Python sys.path, so the modules can be loaded correctly. Then, you will likely want to inherit the same CLI parser defined in lib/options.py to save time. You can then add a new argument group app to the parser to add custom CLI arguments to be used in conjunction with the defaults. See app/sdf_renderer.py for an example.

Examples of things that are considered apps include, but are not limited to:

visualizers
training code
downstream applications

Third-Party Libraries

This code includes code derived from 3 third-party libraries, all distributed under the MIT License:

https://github.com/zekunhao1995/DualSDF

https://github.com/rogersce/cnpy

https://github.com/krrish94/nerf-pytorch

Acknowledgements

We would like to thank Jean-Francois Lafleche, Peter Shirley, Kevin Xie, Jonathan Granskog, Alex Evans, and Alex Bie at NVIDIA for interesting discussions throughout the project. We also thank Peter Shirley, Alexander Majercik, Jacob Munkberg, David Luebke, Jonah Philion and Jun Gao for their help with paper editing.

We also thank Clement Fuji Tsang for his help with the code release.

The structure of this repo was inspired by PIFu: https://github.com/shunsukesaito/PIFu

tovacinni / nglod