Official code for Object Counting and Instance Segmentation with Image-level Supervision, in CVPR 2019 and TPAMI 2020

Paper (conference version) Paper (journal version) [Supp]() [Presentation]() [Poster]()

This code is built by Guolei Sun. For questions, please contact sunguolei.kaust@gmail.com

Requirements:

1. System: ubuntu 16.04.
2. NVIDIA GPU + CUDA CuDNN
3. Python>=3.6
4. Pytorch version 0.4
5. Jupyter Notebook and ipywidgets
6. Other common packages: numpy, scipy, and so on. Please refer to environment.yml.

Installation:

This respority uses some functions from PRM.

1. Download this respority and unzip it. Make sure that the folders look like this:

   CountSeg
   ├── Nest-pytorch
     ├── ...
   ├── PRM-pytorch
     ├── ...
   ├── ...

2. Go inside to "CountSeg" folder by "cd path/CountSeg", where path is where you store CountSeg in your computer.
3. Install Nest, a flexible tool for building and sharing deep learning modules, created by Yanzhao

   pip install git+https://github.com/ZhouYanzhao/Nest.git

4. Install PRM via Nest's CLI tool

   nest module install ./PRM-pytorch prm

   Validate the installation by "nest module list --filter prm", you should see something like this.

   ## Output:
   #
   # 3 Nest modules found.
   #[0] prm.fc_resnet50 (1.0.0)
   #[1] prm.peak_response_mapping (1.0.0)
   #[2] prm.prm_visualize (1.0.0)

   If you get some error, it is because that you miss some packages. Install them and do the validation again until you can get something like above

5. Install Nest's build-in Pytorch modules

   nest module install ./Nest-pytorch pytorch

Dataset

1. For Pascal, download dataset by simply running script in CountSeg/dataset folder.

   cd path/CountSeg/dataset
   chmod +x pascal_download.sh
   ./pascal_download.sh

   Before running the script, do not forget to change the save_path_2007 and save_path_2012 in pascal_download.sh to the path where you want to save pascal 2007 and 2012 dataset.

2. For COCO dataset, please download dataset from COCO

Demo

Go to "CountSeg/demo" and run demo.

Test

To reproduce the results reproted in the paper.

1. Pascal 2007 counting

   cd path/CountSeg
   jupyter notebook

   Then open eval_counting_pascal07.ipynb and simply run each cell inside it. Make sure you change the data path in eval_counting_pascal07.ipynb.

2. COCO 2014 counting. Open jupyter notebook and run eval_counting_coco14.ipynb. Make sure you change the data path in eval_counting_coco14.ipynb.

Training

1. Pascal 2007 and COCO 2014.

   cd path/CountSeg/experiments
   jupyter notebook

   Then open main-pascal.ipynb or main-coco.ipynb to do training on Pascal or COCO, respectively. Make sure you change the data path in config_counting_pascal07.yml and config_counting_coco14.yml.

Citation

If you find the code useful for your research, please cite:

@INPROCEEDINGS{cholakkal_sun2019object,
    author = {Cholakkal, Hisham and Sun, Guolei and Khan, Fahad Shahbaz and Shao, Ling},
    title = {Object Counting and Instance Segmentation with Image-level Supervision},
    booktitle = {CVPR},
    year = {2019}
}

@article{cholakkal_sun2020towards,
  title={Towards Partial Supervision for Generic Object Counting in Natural Scenes},
  author={Cholakkal, Hisham and Sun, Guolei and Khan, Salman and Khan, Fahad Shahbaz and Shao, Ling and Gool, Luc Van},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
  year={2020}
}