DCANet

Demo

You can directly try a demo on: https://modelscope.cn/models/damo/cv_hrnet_crowd-counting_dcanet/summary

Datasets Preparation

Download the datasets ShanghaiTech A, ShanghaiTech B and UCF-QNRF Then generate the density maps via generate_density_map_perfect_names_SHAB_QNRF_NWPU_JHU.py. After that, create a folder named JSTL_large_dataset, and directly copy all the processed data in JSTL_large_dataset.

The tree of the folder should be:

`DATASET` is `SHA`, `SHB` or `QNRF_large`.

-JSTL_large_dataset
   -den
       -test
            -Npy files with the name of DATASET_img_xxx.npy, which logs the info of density maps.
       -train
            -Npy files with the name of DATASET_img_xxx.npy, which logs the info of density maps.
   -ori
       -test_data
            -ground_truth
                 -MAT files with the name of DATASET_img_xxx.mat, which logs the original dot annotations.
            -images
                 -JPG files with the name of DATASET_img_xxx.mat, which logs the original image files.
       -train_data
            -ground_truth
                 -MAT files with the name of DATASET_img_xxx.mat, which logs the original dot annotations.
            -images
                 -JPG files with the name of DATASET_img_xxx.mat, which logs the original image files.

Download the pretrained hrnet model HRNet-W40-C from the link https://github.com/HRNet/HRNet-Image-Classification and put it directly in the root path of the repository.

After doing that, download the pretrained model via

bash download_models.sh

And put the IDK model into folder './output', change the model name in test.sh or test_fast.sh scripts.

Test

sh test.sh

Or if you have two GPUs, then

sh test_fast.sh

Training

As the whole training of our pipeline is a little bit complex. So we put the whole code of the pipeline into different main folders. Follow the steps below:

1. You need to the baseline of DCANet in the folder Phase1_JSTL, JSTL indicates train all the images from the observed domains together.

2. Then you can use Phase2_Get_weight or Phase2_Get_weight_Fast to extract $\Delta MAE$ for each channel for each image. The code in Phase2_Get_weight is the navie extractly implementation directly upon the definition, however, it is quite slow. The code in Phase2_Get_weight_Fast is much faster.

3. After obtaining the indicators, go to Phase_Cal_domain and select the suitable script on your own. Then you can get the domain kernels recorded in the file *.npz.

4. Copy *.npz to Phase3C_guided_training for DDK training. You should copy the baseline model into the folder and load the pretrained weights for further DDK training. You can get the instructions from the files *.sh.

5. After the model is trained, copy the model into Phase3Cb_fast, also copy *.npz to this folder, you can either perform IDK training directly from DCANet$_{base}$ or from DCANet (${\cal L}_D$).

6. You can also train/test WorldExpo or UCF_CC_50 in the folder Phase0_Train_WE_And_Test_WE_UCFCC.

Eval the model with your known images without gt counts

Refer to test_unknown folder.

Mention

It is suggested that you should create soft links of image folder(e.g., JSTL_large_dataset) and hrnet model(i.e., hrnetv2_w40_imagenet_pretrained) in each main folder. Always, read config.py in each main folder if a path error occurs.

Citation

If you find our paper give your any insights, please cite:

@ARTICLE{yan2021towards,
  author={Yan, Zhaoyi and Li, Pengyu and Wang, Biao and Ren, Dongwei and Zuo, Wangmeng},
  journal={IEEE Transactions on Circuits and Systems for Video Technology}, 
  title={Towards Learning Multi-domain Crowd Counting}, 
  year={2021},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TCSVT.2021.3137593}}