Periodic Vibration Gaussian: Dynamic Urban Scene Reconstruction and Real-time Rendering

[Project] [Paper]

Periodic Vibration Gaussian: Dynamic Urban Scene Reconstruction and Real-time Rendering,
Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, Li Zhang
Arxiv preprint

Official implementation of "Periodic Vibration Gaussian: Dynamic Urban Scene Reconstruction and Real-time Rendering".

🛠️ Pipeline

Get started

Environment

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

# Make a conda environment.
conda create --name pvg python=3.9
conda activate pvg

# 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 data.

Waymo dataset

Preprocessed 4 waymo scenes for results in Table 1 of our paper can be downloaded here (optional: corresponding label). Please unzip and put it into data directory.

First prepare the kitti-format Waymo dataset:

# Given the following dataset, we convert it to kitti-format
# data
# └── waymo
#     └── waymo_format
#         └── training
#             └── segment-xxxxxx

# install some optional package
pip install -r requirements-data.txt 

# Convert the waymo dataset to kitti-format
python scripts/waymo_converter.py waymo --root-path ./data/waymo/ --out-dir ./data/waymo/ --workers 128 --extra-tag waymo

Then use the example script scripts/extract_scenes_waymo.py to extract the scenes from the kitti-format Waymo dataset which we employ to extract the scenes listed in StreetSurf.

Following StreetSurf, we use Segformer to extract the sky mask and put them as follows:

data
└── waymo_scenes
    └── sequence_id
        ├── calib
        │   └── frame_id.txt
        ├── image_0{0, 1, 2, 3, 4}
        │   └── frame_id.png
        ├── sky_0{0, 1, 2, 3, 4}
        │   └── frame_id.png
        |── pose
        |   └── frame_id.txt
        └── velodyne
            └── frame_id.bin

We provide an example script scripts/extract_mask_waymo.py to extract the sky mask from the extracted Waymo dataset, follow instructions here to setup the Segformer environment.

KITTI dataset

Preprocessed 3 kitti scenes for results in Table 1 of our paper can be downloaded here. Please unzip and put it into data directory.

Put the KITTI-MOT dataset in data directory. Following StreetSurf, we use Segformer to extract the sky mask and put them as follows:

data
└── kitti_mot
    └── training
        ├── calib
        │   └── sequence_id.txt
        ├── image_0{2, 3}
        │   └── sequence_id
        │       └── frame_id.png
        ├── sky_0{2, 3}
        │   └── sequence_id
        │       └── frame_id.png
        |── oxts
        |   └── sequence_id.txt
        └── velodyne
            └── sequence_id
                └── frame_id.bin

We also provide an example script scripts/extract_mask_kitti.py to extract the sky mask from the KITTI dataset.

Training

# Waymo image reconstruction
CUDA_VISIBLE_DEVICES=0 python train.py \
--config configs/waymo_reconstruction.yaml \
source_path=data/waymo_scenes/0145050 \
model_path=eval_output/waymo_reconstruction/0145050

# Waymo novel view synthesis
CUDA_VISIBLE_DEVICES=0 python train.py \
--config configs/waymo_nvs.yaml \
source_path=data/waymo_scenes/0145050 \
model_path=eval_output/waymo_nvs/0145050

# KITTI image reconstruction
CUDA_VISIBLE_DEVICES=0 python train.py \
--config configs/kitti_reconstruction.yaml \
source_path=data/kitti_mot/training/image_02/0001 \
model_path=eval_output/kitti_reconstruction/0001 \
start_frame=380 end_frame=431

# KITTI novel view synthesis
CUDA_VISIBLE_DEVICES=0 python train.py \
--config configs/kitti_nvs.yaml \
source_path=data/kitti_mot/training/image_02/0001 \
model_path=eval_output/kitti_nvs/0001 \
start_frame=380 end_frame=431

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

Evaluating

You can also use the following command to evaluate.

CUDA_VISIBLE_DEVICES=0 python evaluate.py \
--config configs/kitti_reconstruction.yaml \
source_path=data/kitti_mot/training/image_02/0001 \
model_path=eval_output/kitti_reconstruction/0001 \
start_frame=380 end_frame=431

Automatically removing the dynamics

You can the following command to automatically remove the dynamics, the render results will be saved in {EXPERIMENT_DIR}/separation directory.

CUDA_VISIBLE_DEVICES=1 python separate.py \
--config configs/waymo_reconstruction.yaml \
source_path=data/waymo_scenes/0158150 \
model_path=eval_output/waymo_reconstruction/0158150

🎥 Videos

🎞️ Demo

🎞️ Rendered RGB, Depth and Semantic

0017085.mp4

0124100.mp4

0147030.mp4

0149060.mp4

🎞️ Image Reconstruction on Waymo

Comparison with static methods

comparison_static_0017085.mp4

comparison_static_0147030.mp4

Comparison with dynamic methods

comparison_dynamic_0017085.mp4

comparison_dynamic_0147030.mp4

🎞️ Novel View Synthesis on Waymo

novel.mp4

📜 BibTeX

@article{chen2023periodic,
  title={Periodic Vibration Gaussian: Dynamic Urban Scene Reconstruction and Real-time Rendering},
  author={Chen, Yurui and Gu, Chun and Jiang, Junzhe and Zhu, Xiatian and Zhang, Li},
  journal={arXiv:2311.18561},
  year={2023},
}