Adapting Lightweight Image-based Counting Models for Video Crowd Counting

official codes of CVPR paper `Adapting Lightweight Image-based Counting Models for Video Crowd Counting'

Some notes

1, Our methods attribute the model continuous prediction ability, which supports the subsampling of the training dataset for efficiency and further preventing overfitting. This strategy is shown in the `strider' parameter. In the efficiency experiments, we use the whole training dataset for a fair comparison.

Dataset Preparation

Download the datasets from official sites. Then go to Dataset/preprocessor.py and run the corresponding basic_washing method. For UCSD, FDST, MALL, VENICE, DRONECROWD datasets, we strongly suggest to run the corresponding basic_washing method, because it will complete all the processing. For VSCROWD dataset, please download it and organize it as follows.

Dataset
├── VSCROWD
│   ├── train_data
|   │   ├── train_001
|   |   │   ├── gt_points
|   |   │   │   ├── 000001.npy 
|   |   │   │   ├── 000002.npy
|   |   │   │   └── ... 
|   |   │   ├── 000001.jpg
|   |   │   ├── 000002.jpg
|   |   │   └── ...   
|   |   │        
|   |   ├── train_002
|   |   ├── train_003
|   |   └── ...
|   └── test_data
|       ├── test_001
|       │   ├── gt_points
|       │   │   ├── 000001.npy 
|       │   │   ├── 000002.npy
|       │   │   └── ... 
|       │   ├── 000001.jpg
|       │   ├── 000002.jpg
|       │   └── ... 
|       │        
|       ├── test_002
|       ├── test_003
|       └── ...
├── FDST
├── UCSD
├── MALL
└── ...

Training

To train the model in the paper, run this command:

python train.py <arg1>

arg1: specify the dataset name--UCSD, FDST, VENICE, MALL, DRONECROWD, or VSCROWD.

e.g.

python train.py UCSD

Evaluation

To evaluate my model, run:

python test.py <arg1> <arg2>

arg1: specify the dataset name--UCSD, FDST, VENICE, MALL, DRONECROWD, or VSCROWD.

arg2: the relative path of the pretrained model.

e.g.

python test.py UCSD Model/model_pretrain/ucsd.tar

Pre-trained Models

You can download my pretrained models from https://github.com/wbshu/SR/releases.

Citation

If the codes help you, please cite

@inproceedings{shu2026adapting,
  title={Adapting Lightweight Image-based Counting Models for Video Crowd Counting},
  author={Shu, Weibo and Chan, Antoni B},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={35227--35237},
  year={2026}
}