PlenOctree Volume Rendering

This is a real-time PlenOctree volume renderer written in C++ using OpenGL, constituting part of the code release for:

PlenOctrees for Real Time Rendering of Neural Radiance Fields
Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, Angjoo Kanazawa

https://alexyu.net/plenoctrees

The project has several repositories:

NeRF-SH training and PlenOctree extraction https://github.com/sxyu/plenoctree
PyTorch PlenOctree rendering CUDA extension https://github.com/sxyu/svox

More will be released soon, we are taking a short break now.

Building

Please install a recent version of CMake https://cmake.org

Linux

mkdir build && cd build
cmake ..
make -j12

If you do not have CUDA-capable GPU, pass -DVOLREND_USE_CUDA=OFF after cmake .. to use fragment shader backend, which is also used for the web demo. It is slower and does not support mesh-insertion and dependent features such as lumisphere probe.

The main real-time PlenOctree rendererer volrend and a headless version volrend_headless are built. The latter requires CUDA. There is also an animation maker volrend_anim, which I used to make some of the video animations; don't worry about it unless interested.

You should be able to build the project as long as you have GLFW. On Ubuntu, you will need X-server; you can try sudo apt-get install libgl1-mesa-dev libxi-dev libxinerama-dev libxcursor-dev libxrandr-dev libgl1-mesa-dev libglu1-mesa-dev

macOS

For macOS, we assume you have the homebrew package manager, and no CUDA-capable GPU.

brew install cmake
brew install glfw
mkdir build && cd build
cmake .. -DVOLREND_USE_CUDA=OFF
export LIBRARY_PATH=$LIBRARY_PATH:/usr/local/lib; export CPLUS_INCLUDE_PATH="/usr/local/Cellar/glfw/3.3.4/include"; make -j8

Windows 10

Install Visual Studio (I am using 2019 here). Then

mkdir build && cd build
cmake .. -G"Visual Studio 16 2019"
cmake --build . --config Release

If you do not have CUDA-capable GPU, pass -DVOLREND_USE_CUDA=OFF after cmake .. to use fragment shader backend, which is also used for the web demo. It is slower and does not support mesh-insertion and dependent features such as lumisphere probe.

Dependencies

C++17
OpenGL
- any dependencies of GLFW
libpng-dev (only for writing image in headless mode and saving screenshot)

Optional

CUDA Toolkit, I tried on both 11.0 and 10.2
- Pass -DVOLREND_USE_CUDA=OFF to disable it.

Run

./volrend <name>.npz

See --help for flags.

There is an ImGui window which exposes rendering options as well as interactive features mentioned in the paper + video. For the mesh insertion, only OBJ files optionally with vertex coloring are supported. Texturing mapping is not implemented right now. Some example meshes are in sample_obj, and a program to generate SH meshes (just for fun) is in sample_obj/sh/gen_sh.cpp. Please use meshlab to triangulate other mesh.

Keyboard + Mouse Controls (Desktop GUI)

Left mouse btn + drag: rotate about camera position
Right mouse btn + drag: rotate about origin point (can be moved)
Middle mouse btn + drag: pan camera
Shift + Left mouse btn + drag: pan camera (alt)
Shift + middle mouse btn + drag: pan camera AND move origin point simultaneously
Scroll with wheel: move forward/back in z
WASDQE: move; Shift + WASDQE to move faster
123456: preset world_up directions, sweep through these keys if scene is using different coordinate system.
0: reset the focal length to default, if you messed with it
Z: cycle gizmo operations translate/rotate/scale (only visible if mesh is opened in Manipulate section)
X: toggle gizmo space local/world (applies to mesh translation/rotation)

Lumisphere probe:

IJKLUO: move the lumisphere probe; Hold shift to move faster

Offscreen Rendering

The program volrend_headless allows you to perform offscreen rendering on a server.

Usage: ./volrend_headless tree.npz -i intrinsics.txt pose1 pose2... [-o out_dir]

intrinsics.txt should be a 4x4 intrinsics matrix. pose1, pose2 ... should contain 3x4 or 4x4 c2w pose matrices, or multiple matrices in a 4Nx4 format. Add -r to use OpenCV camera space instead of NeRF.

The following zip file contains intrinsics and pose files for each scene of NeRF-synthetic, https://drive.google.com/file/d/1mI4xl9FXQDm_0TidISkKCp9eyTz40stE/view?usp=sharing

Example to render out images: ./volrend_headless drums/tree.npz -i data/nerf_synthetic/drums/intrinsics.txt data/nerf_synthetic/drums/pose/* -o tree_rend/drums

The PNG writing is a huge bottleneck. Example to compute the FPS: ./volrend_headless drums/tree.npz -i data/nerf_synthetic/drums/intrinsics.txt data/nerf_synthetic/drums/pose/*

See ./volrend_headless --help for more options such as setting rendering options.

Precomputed PlenOctree Files

The full resolution tree files for NeRF-synthetic reported in the paper may be found at: https://drive.google.com/drive/folders/1DIYj-iu3TOHProJVHPIQTjHnmYf80_vC?usp=sharing

The uncompressed NeRF-synthetic files used for the web demo are here: https://drive.google.com/drive/folders/1vGXEjb3yhbClrZH1vLdl2iKtowfinWOg?usp=sharing The compression script used to turn this in to the web version is in scripts/compress_octree.py.

More to come soon.

PyTorch Extension: svox

You can find a (mostly) compatible PlenOctree library called svox, which we use to build the tree; pip install svox.

Code: https://github.com/sxyu/svox
Documentation: https://svox.readthedocs.io

More information to be added soon.

Building the Web Demo

The backend of the web demo is built from the shader version of the C++ source using emscripten. Install emscripten per instructions here: https://emscripten.org/docs/getting_started/downloads.html

Then use

mkdir embuild && cd embuild
emcmake cmake ..
make -j12

The full website should be written to embuild/build. Some CMake scripts even write the html/css/js files. To launch it locally for previewing, you can use the make target:

make serve

Which should launch a server at http://0.0.0.0:8000/

ignacio-rocco / volrend