canev9910/trips

TRIPS: Trilinear Point Splatting for Real-Time Radiance Field Rendering

Linus Franke, Darius Rückert, Laura Fink, Marc Stamminger

Point-based radiance field rendering has demonstrated impressive results for novel view synthesis, offering a compelling blend of rendering quality and computational efficiency. However, also latest approaches in this domain are not without their shortcomings. 3D Gaussian Splatting [Kerbl and Kopanas et al. 2023] struggles when tasked with rendering highly detailed scenes, due to blurring and cloudy artifacts. On the other hand, ADOP [Rückert et al. 2022] can accommodate crisper images, but the neural reconstruction network decreases performance, it grapples with temporal instability and it is unable to effectively address large gaps in the point cloud. In this paper, we present TRIPS (Trilinear Point Splatting), an approach that combines ideas from both Gaussian Splatting and ADOP. The fundamental concept behind our novel technique involves rasterizing points into a screen-space image pyramid, with the selection of the pyramid layer determined by the projected point size. This approach allows rendering arbitrarily large points using a single trilinear write. A lightweight neural network is then used to reconstruct a hole-free image including detail beyond splat resolution. Importantly, our render pipeline is entirely differentiable, allowing for automatic optimization of both point sizes and positions. Our evaluation demonstrate that TRIPS surpasses existing state-of-the-art methods in terms of rendering quality while maintaining a real-time frame rate of 60 frames per second on readily available hardware. This performance extends to challenging scenarios, such as scenes featuring intricate geometry, expansive landscapes, and auto-exposed footage.

[Project Page] [Paper] [Youtube] [Supplemental Data]

Citation

@article{franke2024trips,
    title={TRIPS: Trilinear Point Splatting for Real-Time Radiance Field Rendering},
    author={Linus Franke and Darius R{\"u}ckert and Laura Fink and Marc Stamminger},
    journal={arXiv preprint arXiv:2401.06003},
    year = {2024}
}

Install Requirements

Supported Operating Systems: Ubuntu 22.04, Windows

Nvidia GPU (lowest we tested was an RTX2070)

Supported Compiler: g++-9 (Linux), MSVC (Windows, we used 19.31.31105.0)

Software Requirement: Conda (Anaconda/Miniconda)

Install Instructions Linux

Install Ubuntu Dependancies

sudo apt install git build-essential gcc-9 g++-9

For the viewer, also install:

sudo apt install xorg-dev

(There exists a headless mode without window management meant for training on a cluster, see below)

Clone Repo

git clone [email protected]:lfranke/TRIPS.git
cd TRIPS/
git submodule update --init --recursive --jobs 0

Create Conda Environment

cd TRIPS
./create_environment.sh

Install Pytorch

cd TRIPS
./install_pytorch_precompiled.sh

Install CuDNN

Either download the latest version and add it to the conda environment (where CUDA 11.8 was installed) or install via conda:

conda activate trips
conda install -y -c conda-forge cudnn=8.9.2

For our experiments, we used CuDNN 8.9.5, however the conda installed version (8.9.2) should also work fine.

Compile TRIPS

cd TRIPS

conda activate trips

export CONDA=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
export CC=gcc-9
export CXX=g++-9
export CUDAHOSTCXX=g++-9

mkdir build
cd build

cmake -DCMAKE_PREFIX_PATH="./External/libtorch/;${CONDA}" ..

make -j10

make can take a long time, especially for some CUDA files.

If you get a undefined reference to ...@GLIBCXX_3.4.30' error during linking, most likely your linker fails to resolve the global and conda version of the c++ standard library.

Consider removing the libstdc++ lib from the conda environment:

cd TRIPS

conda activate trips

export CONDA=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}

rm $CONDA_PREFIX/lib/libstdc++.so*

Install Instructions Windows

Software Requirements:

• VS2022

• CUDA 11.8

• Cudnn (copy into 11.8 folder as per install instructions) (we used version 8.9.7)

• conda (we used Anaconda3)

  [Start VS2022 once for CUDA integration setup]

Clone Repo

git clone [email protected]:lfranke/TRIPS.git
cd TRIPS/
git submodule update --init --recursive --jobs 0

Setup Environment

conda update -n base -c defaults conda

conda create -y -n trips python=3.9.7

conda activate trips

conda install -y cmake=3.26.4
conda install -y -c intel mkl=2024.0.0
conda install -y -c intel mkl-static=2024.0.0
conda install openmp=8.0.1 -c conda-forge

Install libtorch:

• Download: https://download.pytorch.org/libtorch/cu116/libtorch-win-shared-with-deps-1.13.1%2Bcu116.zip
• Unzip
• Copy into TRIPS/External

Folder structure should look like:

TRIPS/
    External/
        libtorch/
            bin/
            cmake/
            include/
            lib/
            ...
        saiga/
        ...
    src/
    ...

Compile

cmake -Bbuild -DCMAKE_CUDA_COMPILER="$ENV:CUDA_PATH\bin\nvcc.exe" -DCMAKE_PREFIX_PATH=".\External\libtorch" -DCONDA_P_PATH="$ENV:CONDA_PREFIX" -DCUDA_P_PATH="$ENV:CUDA_PATH" -DCMAKE_BUILD_TYPE=RelWithDebInfo .

cmake --build build --config RelWithDebInfo -j

The last cmake build call can take a lot of time.

Running on pretrained models

Supplemental materials link: https://zenodo.org/records/10606698

After a successful compilation, the best way to get started is to run viewer on the tanks and temples scenes using our pretrained models. First, download the scenes and extract them into scenes/. Now, download the model checkpoints and extract them into experiments/. Your folder structure should look like this:

TRIPS/
    build/
        ...
    experiments/
        checkpoint_train
        checkpoint_playground
        ...
    scenes/
        tt_train/
        tt_playground/
        ...
    ...

Viewer

Your working directory should be the trips root directory.

Linux

Start the viewer with

conda activate trips
export CONDA=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA/lib
./build/bin/viewer --scene_dir scenes/tt_train

Windows

./build/bin/RelWithDebInfo/viewer.exe  --scene_dir scenes/tt_train

The path is different to the Linux path, the compile configuration is added (RelWithDebInfo)!

Viewer Controls

The most important keyboard shortcuts are:

• F1: Switch to 3DView
• F2: Switch to neural view
• F3: Switch to split view (default)
• F4: Switch to point rendering view
• WASD: Move camera
• Center Mouse + Drag: Rotate around camera center
• Left Mouse + Drag: Rotate around world center
• Right click in 3DView: Select camera
• Q: Move camera to selected camera

By default, TRIPS is compiled with a reduced GUI. If you want all GUI buttons present, you can add a -DMINIMAL_GUI=OFF to the first cmake call to compile this in.

Scene Description

• TRIPS uses ADOP's scene format.
• ADOP uses a simple, text-based scene description format.
• To run on your scenes you have to convert them into this format.
• If you have created your scene with COLMAP (like us) you can use the colmap2adop converter.
• More infos on this topic can be found here: scenes/README.md

Training

The pipeline is fitted to your scenes by the train executable. All training parameters are stored in a separate config file. The basic syntax is:

Linux:

./build/bin/train --config configs/train_normalnet.ini

Windows:

./build/bin/RelWithDebInfo/train.exe --config configs/train_normalnet.ini

Make again sure that the working directory is the root. Otherwise, the loss models will not be found.

Two configs are given for the two networks used in the paper: train_normalnet.ini and train_sphericalnet.ini You can override the options in these configs easily via the command line.

./build/bin/train --config configs/train_normalnet.ini --TrainParams.scene_names tt_train --TrainParams.name new_name_for_this_training

For scenes with extensive environments, consider adding an environment map with:

--PipelineParams.enable_environment_map true

If GPU memory is sparse, consider lowering batch_size (standard is 4), inner_batch_size (standard is 4) or train_crop_size (standard is 512) with for example,

--TrainParams.batch_size 1
--TrainParams.inner_batch_size 2
--TrainParams.train_crop_size 256

(however this may impact quality).

By default, every 8th image is removed during training and used as a test image. If you want to change this split, consider overriding which percentage of images should be kept out of training with:

--TrainParams.train_factor 0.1 

default is 0.125 (so 1/8).

Live Viewer during Training

An experimental live viewer is implemented which shows the fitting process during training in an OpenGL window. If headless mode is not required (see below) you can add a -DLIVE_TRAIN_VIEWER=ON to the first cmake call to compile this version in.

Note: This will have an impact on training speed, as intermediate (full) images will we rendered during training.

Headless Mode

If you do not want the viewer application, consider calling cmake with an additional -DHEADLESS=ON. This is usually done for training on remote machines.