VLM-HCS: Plug-and-Play Hierarchical Coresets Selection for Wide-Area Scene Understanding

This repository provides a Hierarchical Coresets Selection (HCS) module that can be plugged into any open-source VLM baselines to select informative regions in a coarse-to-fine manner for scene understanding tasks.

🚀 Live Demo Website

🥇[ACMMM2025] Advancing Complex Wide-Area Scene Understanding with Hierarchical Coresets Selection

👉 Click here to view the project online

Setup

conda create -n vlm-hcs python=3.10 -y
conda activate vlm-hcs
pip install -r requirements.txt

Quickstart (Example)

bash scripts/demo_clip_hcs.sh

Notes

• Backbones are frozen by default; HCS acts as a pre-selection module.
• You can optionally train a tiny MLP scorer (train/train_hcs_scorer.py) to stabilize scores.

Citation

If you find our work and codes useful, please consider citing our paper and star our repository (🥰🎉Thanks!!!):

@misc{wang2025advancing,
      title={Advancing Complex Wide-Area Scene Understanding with Hierarchical Coresets Selection}, 
      author={Jingyao Wang and Yiming Chen and Lingyu Si and Changwen Zheng},
      year={2025},
      eprint={2507.13061},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.13061}, 
}