Awesome

SS-DCNet

This is the repository for SS-DCNet, presented in our paper:

From Open Set to Closed Set: Supervised Spatial Divide-and-Conquer for Object Counting

Haipeng Xiong¹, Hao Lu², Chengxin Liu¹, Liang Liu¹, Chunhua Shen², Zhiguo Cao¹

¹Huazhong University of Science and Technology, China

²The University of Adelaide, Australia

Contributions

• We propose to transform open-set counting into a closed set problem via S-DC. A theoretical analysis of why such a transformation makes sense is also presented;
• We investigate the explicit supervision for S-DC, which leads to a novel SS-DCNet. SS-DCNet is applicable to both regression-based and classification-based counters and can produce visually clear spatial divisions;
• We report state-of-the-art counting performance over 5 challenging datasets with remarkable relative improvements. We also show good transferablity of SS-DCNet via cross-dataset evaluations on crowd counting datasets.

Results

The mean average error (MAE) of SS-DCNet against its previous version, S-DCNet, over 5 challenging counting datasets are reported on the table:

Environment

Please install required packages according to requirements.txt.

Data

Testing data for ShanghaiTech and UCF-QNRF dataset have been preprocessed. You can download the processed dataset from:

ShanghaiTech PartA [Baidu Yun] with code: po1v or [Google Drive]

ShanghaiTech PartB [Baidu Yun] with code: po1v or [Google Drive]

UCF-QNRF [Baidu Yun] with code: po1v or [Google Drive]

Model

Pretrained weights can be downloaded from:

ShanghaiTech PartA [Baidu Yun] with code: weng or [Google Drive]

ShanghaiTech PartB [Baidu Yun] with code: weng or [Google Drive]

UCF-QNRF [Baidu Yun] with code: weng or [Google Drive]

A Quick Demo

1. Download the code, data and model.

2. Organize them into one folder. The final path structure looks like this:

-->The whole project
    -->data
        -->SH_partA
        -->SH_partB
        -->UCF-QNRF_ECCV18
    -->model
        -->SHA
        -->SHB
        -->QNRF
    -->Network
        -->base_Network_module.py
        -->merge_func.py
        -->class_func.py
        -->SSDCNet.py
    -->all_main.py
    -->main_process.py
    -->Val.py
    -->load_data_V2.py
    -->IOtools.py

3. Run the following code to reproduce our results. The MAE will be SHA: 55.571, SHB: 6.645 and QNRF: 81.864 . Have fun:)

   for ShanghaiTech PartA:
   python all_main.py --dataset SHA 
   
   for ShanghaiTech PartB:
   python all_main.py --dataset SHB 
   
   for UCF-QNRF:
   python all_main.py --dataset QNRF 
   

References

If you find this work or code useful for your research, please cite:

@misc{xiong2020open,
    title={From Open Set to Closed Set: Supervised Spatial Divide-and-Conquer for Object Counting},
    author={Haipeng Xiong and Hao Lu and Chengxin Liu and Liang Liu and Chunhua Shen and Zhiguo Cao},
    year={2020},
    eprint={2001.01886},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

and

@inproceedings{xhp2019SDCNet,
  title={From Open Set to Closed Set: Counting Objects by Spatial Divide-and-Conquer},
  author={Xiong, Haipeng and Lu, Hao and Liu, Chengxin and Liang, Liu and Cao, Zhiguo and Shen, Chunhua},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2019},
  pages = {8362-8371}
}