This is our PyTorch implementation for the Neural-Best Buddies paper.

The code was written by Kfir Aberman and supported by Mingyi Shi.

Neural Best-Buddies: Project | Paper

If you use this code for your research, please cite:

Neural Best-Buddies: Sparse Cross-Domain Correspondence Kfir Aberman, Jing Liao, Mingyi Shi, Dani Lischinski, Baoquan Chen, Daniel Cohen-Or, SIGGRAPH 2018.

• Linux or macOS
• Python 2 or 3
• CPU or NVIDIA GPU + CUDA CuDNN

• Run the algorithm (demo example)

#!./script.sh
python3 main.py --datarootA ./images/original_A.png --datarootB ./images/original_B.png --name lion_cat --k_final 10

The option --k_final dictates the final number of returned points. The results will be saved at ../results/. Use --results_dir {directory_path_to_save_result} to specify the results directory.

Sparse correspondence:

• correspondence_A.txt, correspondence_B.txt
• correspondence_A_top_k.txt, correspondence_B_top_k.txt

Dense correspondence (densifying based on MLS):

• BtoA.npy, AtoB.npy

Warped images (aligned to their middle geometry):

• warp_AtoM.png, warp_BtoM.png

• If you are running the algorithm on a bunch of pairs, we recommend to stop it at the second layer to reduce runtime (comes at the expense of accuracy), use the option --fast.
• If the images are very similar (e.g, two frames extracted from a video), many corresponding points might be found, resulting in long runtime. In this case we suggest to limit the number of corresponding points per level by setting --k_per_level 20 (or any other desired number)

If you use this code for your research, please cite our paper:

@article{aberman2018neural,
  title={Neural best-buddies: Sparse cross-domain correspondence},
  author={Aberman, Kfir and Liao, Jing and Shi, Mingyi and Lischinski, Dani and Chen, Baoquan and Cohen-Or, Daniel},
  journal={ACM Transactions on Graphics (TOG)},
  volume={37},
  number={4},
  pages={69},
  year={2018},
  publisher={ACM}
}