[ICCV 2025] CityGS-X : A Scalable Architecture for Efficient and Geometrically Accurate Large-Scale Scene Reconstruction

Northwestern Polytechnical University; Shanghai Artificial Intelligence Laboratory

Yuanyuan Gao*, Hao Li*, Jiaqi Chen*, Zhengyu Zou, Zhihang Zhong†, Dingwen Zhang†, Xiao Sun, Junwei Han
(* indicates equal contribution, † means co-corresponding author)

This repo contains official implementations of CityGS-X, ⭐ us if you like it!

• 🔥🔥 News: 2025/4/17: training & inference code is now available! You can try it.
• 🔥🔥 News: 2025/6/28: CityGS-X has been accepted to ICCV 2025.

• Release the training & inference code of CityGS-X.
• Reproduction scripts, records, and results under different parameter settings are available at yy456/CityGS-X. Many thanks to @gongshiqiang02 for providing the reproduction results.

We tested CityGS-X on a server configured with Ubuntu 18.04, cuda 11.6 and gcc 9.4.0. Other similar configurations should also work, but we have not verified each one individually.

1. Clone this repo:

git clone https://github.com/gyy456/CityGS-X.git --recursive
cd CityGS-X

2. Install dependencies

SET DISTUTILS_USE_SDK=1 # Windows only
conda env create --file environment.yml
conda activate citygx-x
pip install submodule_cityx/diff-gaussian-rasterization
pip install submodule_cityx/simple-knn

When training on a synthetic dataset, depth maps can be produced and they do not require further processing to be used in our method.

For real world datasets depth maps should be generated for each input images, to generate them please do the following:

1. Clone Depth Anything v2 （You can try other depth estimation models）:

   git clone https://github.com/DepthAnything/Depth-Anything-V2.git
   

2. Download weights from Depth-Anything-V2-Large and place it under Depth-Anything-V2/checkpoints/
3. Generate depth maps (set the depth image reslution align with the training reslution you want):

   python Depth-Anything-V2/run.py --encoder vitl --pred-only --grayscale \
   --img-path <path to input images> --outdir <output path>
   

4. Generate a depth_params.json file using:

   python utils/make_depth_scale.py --base_dir <path to colmap> --depths_dir <path to generated depths>
   

5. use the multi-view constrains to filter the depth:

   python  multi_view_precess.py  -s  datasets/<scene_name> --resolution 4 \
   --model_path datasets/<scene_name>/train/mask  --images train/rgbs  --pixel_thred 1
   

• pixel_thred: set the threshold of the pixel position loss;

First, create a data/ folder inside the project path by

mkdir data

Download the datasets following the Mega-NeRF repository.

After downloading, for Mill-19 and UrbanScene3D, run the following code for each dataset:

python tools/merge_val_train.py -d $DATASET_DIR(data/<scene_name>)

bash tools/colmap_full.sh  $COLMAP_RESULTS_DIR  $DATASET_ROOT(data/<scene_name>)

While for the MatrixCity, CityGS-X follow the preprocess of CityGaussianV2

The data structure will be organised as follows:

data/
├── scene_name(Mill-19 and UrbanScene3D)
│   ├── train/
│   │   ├── rgbs
│   │   │   ├── 000000.jpg
│   │   │   ├── 000001.jpg
│   │   │   ├── ...
│   │   ├── depths
│   │   │   ├── 000000.png
│   │   │   ├── 000001.png
│   │   ├── mask
│   │   │   ├── 000000.png
│   │   │   ├── 000001.png
│   │   │   ├── ...
│   ├── val/
│   │   ├── rgbs
│   │   │   ├── 000000.jpg
│   │   │   ├── 000001.jpg
│   │   │   ├── ...
│   ├── sparse/
│   │   └──0/
├── scene_name(MatrixCity)
│   ├── train/block_all
│   │   ├── images
│   │   │   ├── 0000.png
│   │   │   ├── 0001.png
│   │   │   ├── ...
│   │   ├── depth
│   │   │   ├── 0000.png
│   │   │   ├── 0001.png
│   │   │   ├── ...
│   │   ├── mask
│   │   │   ├── 0000.png
│   │   │   ├── 0001.png
│   │   │   ├── ...
│   │   ├── sparse
│   │       └──0/
│   ├── test/block_all_test
│   │   ├── images
│   │   │   ├── 0000.png
│   │   │   ├── 0001.png
│   │   │   ├── ...
│   │   ├── sparse/
│   │       └──0/
...

To train multiple scenes in parallel, we provide batch training scripts:

• Mill-19 and UrbanScene3D: train_mill19.sh

• MatrixCity: train_matrix_city.sh

run them with

bash train_xxx.sh

• not_use_dpt_loss: you can jump Step2 depth supervision;
• not_use_multi_view_loss: you can jump Step3 multi-view geometric constrains;
• not_use_single_view_loss: you can choose not use the single-view geometric loss;
• gpu_num: specify the GPU number to use;
• bsz: set the taining batch size;
• iteration: set the whole training iterations;
• single_view_weight_from_iter: set the start iteration of the single-view geometric loss default 10_000;
• scale_loss_from_iter: set the start iteration of the scale loss default 0;
• dpt_loss_from_iter: set the start iteration of the depth supervision default 10_000;
• multi_view_weight_from_iter: seet the start iteration of multi-view constrains default 30_000;
• default_voxel_size: set the mean voxel size of the anchor default 0.001; (default_voxel_size will influence the final anchors number)
• distributed_dataset_storage: if cpu memory is enough set it False (Load all the RGB depth and gray image on every process), if cpu memory is not enough， set it Ture (Load RGB depth and gray image on one process and broadcast to other process).
• distributed_save: if Ture load the final model seperately by process, if False load the final model in one model(default)
• default_voxel_size: set the default voxel size for initialization.
• dpt_end_iter: set the end iteration of step2 depth supervision.
• multi_view_patch_size: the multi-view patchh loss is calculated by gray image, for the less texture scene or higher reslution, larger patch_size works better, but may caused longer training time.

torchrun --standalone --nnodes=1 --nproc-per-node=<gpu_num>  train.py --bsz <bsz> -s datasets/<scene_name> \
    --resolution 4 --model_path output/<save_path> --iterations 100000 --images train/rgbs \
    --single_view_weight_from_iter 10000  --depth_l1_weight_final 0.01 --depth_l1_weight_init 0.5 \
    --dpt_loss_from_iter 10000  --multi_view_weight_from_iter 30000 --default_voxel_size 0.001 \
    --dpt_end_iter 30_000 --multi_view_patch_size 3

python train.py --bsz <bsz> -s datasets/<scene_name> --resolution 4 --model_path output/<save_path> \
    --iterations 100000 --images train/rgbs --single_view_weight_from_iter 10000 \
    --depth_l1_weight_final 0.01 --depth_l1_weight_init 0.5 --dpt_loss_from_iter 10000 \
    --multi_view_weight_from_iter 30000 --default_voxel_size 0.001 --dpt_end_iter 30000 \
    --multi_view_patch_size 3

The training time may faster than the table provided in our paper, as we have optimize the multi-process dataloader.

Evalutaion image is saved and PSNR is calcuated during training by default except MartrixCity.

torchrun --standalone --nnodes=1 --nproc-per-node=<gpu_num>  render.py --bsz <bsz> \
    -s datasets/<scene_name> --resolution 4 --model_path output/<save_path> \
    --images train/rgbs --skip_train

python render.py --bsz <bsz> -s datasets/<scene_name> --resolution 4 \
    --model_path output/<save_path> --images train/rgbs --skip_train

python metrics.py -m output/<save_path>

torchrun --standalone --nnodes=1 --nproc-per-node=<gpu_num>  render_mesh.py --bsz <bsz> \
    -s datasets/<scene_name> --resolution 4 --model_path output/<save_path> \
    --images train/rgbs --voxel_size 0.001 --max_depth 5 --use_depth_filter

python render_mesh.py --bsz <bsz> -s datasets/<scene_name> --resolution 4 \
    --model_path output/<save_path> --images train/rgbs --voxel_size 0.001 \
    --max_depth 5 --use_depth_filter

• voxel_size: set the mesh voxel size.

python eval_f1.py --ply_path_pred <mesh_path> --ply_path_gt <gt_point_cloud_path> --dtau 0.5

We would like to express our gratitude to the authors of the following algorithms and libraries, which have greatly inspired and supported this project:

• Grendel-GS: On Scaling Up 3D Gaussian Splatting Training
• PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction
• Octree-GS: Towards Consistent Real-time Rendering with LOD-Structured 3D Gaussians
• CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes
• Momentum-GS: Momentum Gaussian Self-Distillation for High-Quality Large Scene Reconstruction

Your contributions to the open-source community have been invaluable and are deeply appreciated.

@misc{gao2025citygsxscalablearchitectureefficient,
      title={CityGS-X: A Scalable Architecture for Efficient and Geometrically Accurate Large-Scale Scene Reconstruction}, 
      author={Yuanyuan Gao and Hao Li and Jiaqi Chen and Zhengyu Zou and Zhihang Zhong and Dingwen Zhang and Xiao Sun and Junwei Han},
      year={2025},
      eprint={2503.23044},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.23044}, 
}