README.md

Neural SDF Flow for 3D Reconstruction of Dynamic Scenes

This is the code for the paper

Wei Mao, Richard Hartley, Mathieu Salzmann, Miaomiao Liu. Neural SDF Flow for 3D Reconstruction of Dynamic Scenes. In ICLR 24.

[paper] [talk]

Dependencies

• Python >= 3.8
• Open3D >= 0.15.0
• PyTorch >= 2.0
• Tensorboard

Datasets

CMU Panoptic dataset

For the original dataset please visit their website http://domedb.perception.cs.cmu.edu/.

Here we provide processed clips from five videos of the dataset: band1, cell1, hanggling_b2, ian3 and pizza1. Download the data.zip and unzip it to the root folder of the project.

Training and Evaluation

• The training and evaluation script is included in ./code/run.sh file.

Citing

If you find our code usefull, please also consider to cite our work

@inproceedings{mao2024neural,
  title={Neural SDF Flow for 3D Reconstruction of Dynamic Scenes},
  author={Mao, Wei and Hartley, Richard and Salzmann, Mathieu and Liu, Miaomiao},
  journal={ICLR},
  year={2024}
}

Acknowledgments

The overall code framework (dataloading, training, testing etc.) is adapted from VolSDF.

The auto computation of gradient in forward call is adapted from DNS

Licence

MIT