DENSER: 3D Gaussians Splatting for Scene Reconstruction of Dynamic Urban Environments

Scene Reconstruction Results

Ground Truth

Reconstruction

Installation

Installing DENSER:

Create environment

conda create --name denser -y python=3.8
conda activate denser
pip install --upgrade pip

Clone the Repository

git clone https://github.com/sntubix/denser.git --recursive

Install Dependencies

cd denser
bash installation.sh

Install denser

pip install -e .

Data Organization

The KITTI-MOT dataset should be organized as follows:

.(KITTI_MOT_ROOT)
└── training
    ├── calib
    │   └── sequence_id.txt
    ├── completion_02                # (Optional) depth completion
    │   └── sequence_id
    │       └── frame_id.png
    ├── completion_03                # (Optional) depth completion
    │   └── sequence_id
    │       └── frame_id.png
    ├── image_02
    │   └── sequence_id
    │       └── frame_id.png
    ├── image_03
    │   └── sequence_id
    │       └── frame_id.png
    ├── label_02
    │   └── sequence_id.txt
    ├── object_lidars
    │   └── object_id.ply
            ....
    └── oxts
        └── sequence_id.txt

Training,Rendering,Evaluation

ds-train denser --data /data/kitti/image_02/0006'
ds-render --load_config /path/to/your/config/config.yml
ds-eval --load_config /path/to/your/config/config.yml

preprocessed Datasets

coming soon

pretrained models

coming soon

Citation

@misc{mohamad2024denser,
      title={DENSER: 3D Gaussians Splatting for Scene Reconstruction of Dynamic Urban Environments}, 
      author={Mahmud A. Mohamad and Gamal Elghazaly and Arthur Hubert and Raphael Frank},
      year={2024},
      eprint={2409.10041},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2409.10041}, 
}