4DGF.md

Project Page | Paper | ArXiv

Before running the model, please follow the instructions in README.md to download and preprocess the data.

Training

To reproduce the experiments in the paper, use the following commands

Argoverse2

ns-train 4dgf-av2-big --machine.num-devices 8 --pipeline.model.max-num-gaussians 8000000 --pipeline.model.object-grid-log2-hashmap-size 17 street --data data/Argoverse2/metadata_PIT_6180_1620_6310_1780.pkl --voxel-size 0.15
ns-train 4dgf-av2-big --machine.num-devices 8 --pipeline.model.max-num-gaussians 8000000 --pipeline.model.object-grid-log2-hashmap-size 17 street --data data/Argoverse2/metadata_PIT_1100_-50_1220_150.pkl --voxel-size 0.15

Waymo

ns-train 4dgf-waymo --data data/waymo/metadata_<sequence_name>.pkl

VKITTI2

ns-train 4dgf-vkitti street --data data/VKITTI2/metadata_02.pkl --train-split-fraction [0.75|0.5|0.25]
ns-train 4dgf-vkitti street --data data/VKITTI2/metadata_06.pkl --train-split-fraction [0.75|0.5|0.25]
ns-train 4dgf-vkitti street --data data/VKITTI2/metadata_18.pkl --train-split-fraction [0.75|0.5|0.25]

KITTI

ns-train 4dgf-kitti street --data data/KITTI/tracking/training/metadata_0001.pkl --train-split-fraction [0.75|0.5|0.25]
ns-train 4dgf-kitti street --data data/KITTI/tracking/training/metadata_0002.pkl --train-split-fraction [0.75|0.5|0.25]
ns-train 4dgf-kitti street --data data/KITTI/tracking/training/metadata_0006.pkl --train-split-fraction [0.75|0.5|0.25]

Citation

@InProceedings{fischer2024dynamic,
    author    = {Tobias Fischer and Jonas Kulhanek and Samuel Rota Bul{\`o} and Lorenzo Porzi and Marc Pollefeys and Peter Kontschieder},
    title     = {Dynamic 3D Gaussian Fields for Urban Areas},
    booktitle = {The Thirty-eighth Annual Conference on Neural Information Processing Systems},
    year      = {2024}
}