1. Environment

Follow the original instructions,

conda create --name street-gaussians-ns -y python=3.8
conda activate street-gaussians-ns
pip install --upgrade pippip install torch==2.1.2+cu118 torchvision==0.16.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118

Bug 1

and when running this line:

conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit

Some error messages:

CondaVerificationError: The package for libcufft-dev located at /root/miniconda3/pkgs/libcufft-dev-10.9.0.58-0
appears to be corrupted. The path 'include/cufftw.h'
specified in the package manifest cannot be found.SafetyError: The package for libcufft-dev located at /root/miniconda3/pkgs/libcufft-dev-10.9.0.58-0
appears to be corrupted. The path 'lib/libcufft_static_nocallback.a'
has an incorrect size.reported size: 308031378 bytesactual size: 299421696 bytesClobberError: This transaction has incompatible packages due to a shared path.packages: nvidia/label/cuda-11.8.0/linux-64::libcusparse-11.7.5.86-0, nvidia/label/cuda-11.8.0/linux-64::libcusparse-dev-11.7.5.86-0path: 'lib/libcusparse.so.11'

Solutions: Use a different mirror source

conda config --add channels defaults
conda install -c defaults cudatoolkit=11.8

Then follow:

pip install ninja git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch
pip install nerfstudio==1.0.0git clone https://github.com/LightwheelAI/street-gaussians-ns.git --recursivecd street-gaussians-nspip install -e .pip install dependencies/nvdiffrast/# Install data processsing tools
cd dependencies/detectron2
pip install -e .cd dependencies/Mask2Former
pip install -r requirements.txt
cd mask2former/modeling/pixel_decoder/ops
sh make.sh

2. Training

We directly use the preprocessed dataset. When training the model

bash scripts/shells/train.sh /root/autodl-tmp/8398516118967750070_3958_000_3978_000 0

Bug 2

  File "/root/autodl-fs/Projects/street-gaussians-ns/street_gaussians_ns/data/sgn_dataparser.py", line 478, in _load_3D_pointsassert points_filepath.exists()
AssertionError

This is because we don’t have the points3D_withlidar.txt file

Solution1:

Generate with scripts/shells/points_cloud_generate.sh

Solution2:

Change to --init_points_filename points3D.bin

Bug 3

project_gaussians() missing 1 required positional argument: 'tile_bounds'

Solution:

Use gsplat with version >=0.1.9 : pip install gsplat==0.1.9

Final visualization:

完结撒花★,°:.☆(￣▽￣)/$:.°★ 。