Could you please release a data preparation guide?

[Six-Bit-TX]

Thank you for your amazing work!

I was wondering if you could release a data preparation guide (for Matrix City in particular). How should I preprocess the data, and how should I organize the data directory?

Your help is much appreciated!

[zhouilu]

We use Gaussian point cloud or dense point cloud for training. For the Matrix City dataset, we use dense reconstruction to obtain dense point cloud, use the result as the initial point cloud, and start training.

The Matrix City dataset can be found here.
https://github.com/Linketic/CityGaussian/blob/main/doc/data_preparation.md

We will update the data preparation guide later.

[zhouilu]

data download

• Download small_city version of aerial data of MatricCity from https://github.com/Linketic/CityGaussian/blob/main/doc/data_preparation.md
• Unzip the data. You can use scripts from https://github.com/Linketic/CityGaussian/blob/main/scripts/untar_matrixcity_train.sh

reconstruction

• Sparse reconstruction with COLMAP, you can use scripts from https://github.com/Linketic/CityGaussian/blob/main/tools/convert_cam.py
• Run following command to get dense reconstruction result

COLMAP_EXE_PATH="/home/admin/bin/"
$COLMAP_EXE_PATH/colmap patch_match_stereo \
  --workspace_path . \
  --workspace_format COLMAP \
  --PatchMatchStereo.max_image_size 2000 \
  --PatchMatchStereo.geom_consistency true
$COLMAP_EXE_PATH/colmap stereo_fusion \
  --workspace_path . \
  --workspace_format COLMAP \
  --input_type geometric \
  --output_path ./fused.ply

train

pcd_path in https://github.com/GaodeRender/triangle-splatting/blob/main/config/MatrixCity_VanillaTS_mesh.yaml should be set to path of fused.ply, then you ran run command python run_experiments.py --type MatrixCity_mesh --dataset_path /path/to/dataset

This guide can be used as a reference. If you have any questions, please ask