【FasterGS】Unbuntu22.04搭建FasterGS（在gaussian-splatting中使用其加速）

文章目录前言一、环境搭建1.1 CUDA12.1安装1.2 虚拟环境构建1.2.1 创建环境1.2.2 安装simple-knn1.2.3 安装fused-ssim1.2.4 安装核心光栅化模块diff-gaussian-rasterization_fastgs1.3 安装 Faster-GS CUDA backend二、测试2.1 训练2.2 渲染2.3 计算评估指标前言主要用于自己搭建环境。我采用Python 3.10, PyTorch 2.1.0, CUDA 12.1来构建虚拟环境。将FasterGS加速用在原版gaussian-splatting上。代码地址FasterGS以及修改后的gaussian-splatting一、环境搭建下载项目文件并进入文件夹gitclone https://github.com/fhahlbohm/gaussian-splatting.git--recursivecdgaussian-splatting1.1 CUDA12.1安装如何安装CUDA可参考Unbuntu22.04中安装多版本的CUDA可任意切换版本1.2 虚拟环境构建1.2.1 创建环境首先将environment_cuda12.yml文件中的内容改为name: gaussian_splatting121 channels: - pytorch - nvidia - conda-forge - defaults dependencies: -python3.10- pip - plyfile - tqdm - opencv - joblib -pytorch2.1.0 -torchvision0.16.0 -torchaudio2.1.0 - pytorch-cuda12.1然后condaenvcreate--fileenvironment_cuda12.yml conda activate gaussian_splatting1211.2.2 安装simple-knnpipinstall--no-build-isolation ./submodules/simple-knn报错1setuptools 没装 / 版本坏了pip uninstall -y setuptools然后pip install setuptools68.0.0最后重新pip install --no-build-isolation ./submodules/simple-knn1.2.3 安装fused-ssimpipinstall--no-build-isolation ./submodules/fused-ssim1.2.4 安装核心光栅化模块diff-gaussian-rasterization_fastgspipinstall--no-build-isolation ./submodules/diff-gaussian-rasterization报错1.需要降级numpypip install numpy1.26.4以及pip install colorama pytz清理旧编译缓存cdsubmodules/diff-gaussian-rasterizationrm-rfbuild *.egg-infocd../..2.rasterizer子模块版本冲突cdsubmodulesrm-rfdiff-gaussian-rasterizationgitclone https://github.com/graphdeco-inria/diff-gaussian-rasterizationcd..重新编译pip install --no-build-isolation ./submodules/diff-gaussian-rasterization1.3 安装 Faster-GS CUDA backendpipinstallgithttps://github.com/nerficg-project/faster-gaussian-splatting/#subdirectoryFasterGSCudaBackend --no-build-isolation二、测试2.1 训练在终端进行位于gaussian_splatting中。下载数据集TanksTemples and Deep Blending我将数据集放在gaussian_splatting文件夹下创建一个datasets文件夹中。在gaussian_splatting文件夹下打开终端conda activate gaussian_splatting121 python train.py-s./datasets/tandt_db/tandt/truck--eval--disable_viewer出现训练结果会保存到output文件夹下保存了7000次迭代我就打断了训练。与原GS方法对比时间花费50s而原GS模型在使用sparse_adam 优化器的情况下花费了151s训练速度提高约2.2倍。文件大小409.8MBFastGs 417.9MBGS几乎没变用3DGS中的3. SIBR查看器可查看结果2.2 渲染python render.py-m./output/ff06f445-f出现渲染结果位于2.3 计算评估指标python metrics.py-m./output/ff06f445-f结果原GS为相差不大比原GS略差。