Optimal Transport Minimization

Official code for CVPR-2023 paper "Optimal Transport Minimization: Crowd Localization on Density Maps for Semi-Supervised Counting"

• [Jun-21-2023] Update the OT-M algorithm and corresponding demo (demo.ipynb).
• The semi-supervised counting framework is preparing. I will release it as soon as possible.

OT-M for Localization

OT-M algorithm is presented in the otm.py. The demo is shown in demo.ipynb:

from otm import den2seq
den2seq(denmap, scale_factor=8, max_itern=16, ot_scaling=0.75)

• denmap is the density map with the shape of [H, W];
• scale_factor means the resolution ratio of image and density map (here it is 8 since the density map is 1/8 of input image);
• max_itern means the maximum numper of OT and M step;
• ot_scaling controls the step number of Sinkhorn algorithm, it is a value in (0, 1). ot_scaling $\to 1$ results in more iterations in OT step.

Citation

@InProceedings{Lin_2023_CVPR,
    author    = {Lin, Wei and Chan, Antoni B.},
    title     = {Optimal Transport Minimization: Crowd Localization on Density Maps for Semi-Supervised Counting},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {21663-21673}
}