MovingDroneCrowd-Weights

This repository provides pretrained weights for MovingDroneCrowd and MovingDroneCrowd++, as presented in the following works:

• Extended version: Video Individual Counting and Tracking from Moving Drones: A Benchmark and Methods
• Conference version: Video Individual Counting for Moving Drones (ICCV 2025 Highlight)

The repository includes models for video individual counting and tracking from moving drones under diverse aerial conditions.

Links

• Code: https://github.com/fyw1999/MovingDroneCrowd
• Dataset: fyw1999/MovingDroneCrowd

Methods

• SDNet: A video individual counting method that predicts shared density maps between adjacent frames via cross-frame attention.
• GD³A: Global Density map Decomposition via group-wise Descriptor Association, providing a more interpretable and efficient video individual counting method.
• DVTrack: A pedestrian tracking method (Descriptor Voting Track) that converts descriptor-level matching into instance-level associations.

Usage

You can download all pretrained weights using the huggingface_hub library:

from huggingface_hub import snapshot_download

snapshot_download(repo_id="fyw1999/MovingDroneCrowd-Weights")

For detailed instructions on training and evaluation, please refer to the official GitHub repository.

Citation

If this project helps your research, please cite the following papers:

@article{MDC++_GD3A,
  title={Video Individual Counting and Tracking from Moving Drones: A Benchmark and Methods},
  author={Fan, Yaowu and Wan, Jia and Han, Tao and Ma, Andy J. and Ouyang, Wanli and Chan, Antoni B.},
  journal={arXiv preprint arXiv:2601.12500},
  year={2026}
}

@inproceedings{MDC_SDNet,
  title={Video Individual Counting for Moving Drones},
  author={Fan, Yaowu and Wan, Jia Friend and Han, Tao and Chan, Antoni B. and Ma, Andy J.},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  month={October},
  year={2025},
  pages={12284--12293}
}