mvp18 / sp2-360

Code for Sp2360: Sparse-view 360° Scene Reconstruction using Cascaded 2D Diffusion Priors

Sp²360: Sparse-view 360° Scene Reconstruction using Cascaded 2D Diffusion Priors

Soumava Paul · Christopher Wewer · Bernt Schiele · Jan Eric Lenssen

Arxiv preprint

Dataset

Our experiments are based on the MipNeRF360 benchmark available here and here. Download and put the dataset under data/360_v2 in the root directory.

Implementation

Currently the repository only has functionalities for our view selection algorithm outlined in A.1. The full implementation of the method will be released soon.

To sample an M-view subset from a given multiview set, run:

python -m utils.view_splits --source_path data/360_v2/<scene>/<scene>_train --test_cams data/360_v2/<scene>/<scene>_test --num_eval_views M  --eval

where <scene> ∈ {bicycle, garden, flowers, stump, treehill, room, counter, kitchen, bonsai}.

🎓 Cite

If this code was helpful for your research, consider citing:

@article{paul2024sp2360,
          title={Sp2360: Sparse-view 360 Scene Reconstruction using Cascaded 2D Diffusion Priors},
          author={Paul, Soumava and Wewer, Christopher and Schiele, Bernt and Lenssen, Jan Eric},
          journal={arXiv preprint arXiv:2405.16517},
          year={2024}
}        

sp2-360's Issues

About training view selection

Thanks for your nice project.
I want to select sparse training views in some complex scene such as mipnerf360, scannet or tanks&temples so that these views can cover most of the whole scene with less overlap between views. So could you please provide the code for data selection in your project?
Thanks a lot.