BézierGS: Dynamic Urban Scene Reconstruction with Bézier Curve Gaussian Splatting

[Project] [Paper]

BézierGS: Dynamic Urban Scene Reconstruction with Bézier Curve Gaussian Splatting
Zipei Ma^⚖, Junzhe Jiang^⚖, Yurui Chen, Li Zhang^✉
Shanghai Innovation Institute; School of Data Science, Fudan University
ICCV 2025

Official implementation of "BézierGS: Dynamic Urban Scene Reconstruction with Bézier Curve Gaussian Splatting".

🛠️ Pipeline

🎞️ Demo

BézierGS.mp4

pedestrian.mp4

🚀 Get started

Environment

# Clone the repo.
git clone https://github.com/fudan-zvg/BezierGS
cd BezierGS

# Make a conda environment.
conda create --name bezier python=3.10 -y
conda activate bezier

# Install requirements.
pip install -r requirements.txt

# Install simple-knn
git clone https://gitlab.inria.fr/bkerbl/simple-knn.git
pip install ./simple-knn

# a modified gaussian splatting (for feature rendering)
git clone --recursive https://github.com/SuLvXiangXin/diff-gaussian-rasterization
pip install ./diff-gaussian-rasterization

# Install nvdiffrast (for Envlight)
git clone https://github.com/NVlabs/nvdiffrast
pip install ./nvdiffrast

Data preparation

Create a directory for the data: mkdir dataset. We provide some processed data here.

Prepare Waymo Open Dataset.

We provide the split file following EmerNeRF. You can refer to this document for download details.

Preprocess the data

Preprocess the example scenes

python script/waymo/waymo_converter.py --root_dir TRAINING_SET_DIR --save_dir SAVE_DIR --split_file script/waymo/waymo_splits/demo.txt --segment_file script/waymo/waymo_splits/segment_list_train.txt

Generating LiDAR depth

python script/waymo/generate_lidar_depth.py --datadir DATA_DIR

Generating sky mask

Install GroundingDINO following this repo and download SAM checkpoint from this link.

python script/waymo/generate_sky_mask.py --datadir DATA_DIR --sam_checkpoint SAM_CHECKPOINT

Generating intance segmentation

git clone https://github.com/xiao10ma/Grounded-Segment-Anything.git
cd Grounded-Segment-Anything

follow the instruction in the repo to install the dependencies.

Run the following command to generate the instance segmentation.

bash waymo_run.sh

Training

CUDA_VISIBLE_DEVICES=0 python train.py \
--config configs/waymo/017.yaml \
source_path=dataset/017 \
model_path=eval_output/waymo_nvs/017

After training, evaluation results can be found in {EXPERIMENT_DIR}/eval_output directory.

Evaluating

You can also use the following command to evaluate.

CUDA_VISIBLE_DEVICES=0 python evaluate.py \
--config configs/waymo/017.yaml \
source_path=dataset/017 \
model_path=eval_output/waymo_nvs/017 \
checkpoint=eval_output/waymo_nvs/017/chkpnt30000.pth

📜 BibTeX

@inproceedings{Ma2025BezierGS,
  title={BézierGS: Dynamic Urban Scene Reconstruction with Bézier Curve Gaussian Splatting},
  author={Ma, Zipei and Jiang, Junzhe and Chen, Yurui and Zhang, Li},
  booktitle={ICCV},
  year={2025},
}