DKPNet

ICCV 2021 Variational Attention: Propagating Domain-Specific Knowledge for Multi-Domain Learning in Crowd Counting

Baseline of DKPNet is available.

Currently, only code of DKPNet-baseline is released.

MSE vs RMSE

In fact, MSE in our paper is equivalent to RMSE in academic papers. Please use the word RMSE instead of MSE when refering to the corresponding numerical values in our paper. We are sorry for the mistake and can do nothing to corret it after the camera-ready version deadline.

Datasets Preparation

Download the datasets ShanghaiTech A, ShanghaiTech B, UCF-QNRF and NWPU Then generate the density maps via generate_density_map_perfect_names_SHAB_QNRF_NWPU_JHU.py. After that, create a folder named JSTL_large_4_dataset, and directly copy all the processed data in JSTL_large_4_dataset.

The tree of the folder should be:

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

Train

sh run_JSTL.sh

Training notes

There are two types of training scripts: train_fast and train_slow. The main differences between them exist in the evaluation procedure. In train_slow, the test images are processed in the main GPU, making the whole training very slow. As the sizes of test images vary largely with each other (the maximum size / the minimun size equals up to 5x !), making the batch size of evaluation can only be 1 on a single GPU. From our observation, the bottleneck lies in the evaluation stage (Maybe 10x computation time longer than the training time), it is not meaningful enough if you train the whole dataset with more GPUs as long as the evaluation processing is still on a single GPU. To this end, we manage to evaluate two images on two GPUs at the same time, as what train_fast does. We think two GPUs are enough for training the whole dataset in the affordable time (~2 days).

It is notable that the batch size of training should be no smaller than 32, or the performance may degrade to some extent.

Test

Download the pretrained model via

bash download_models.sh

And put the model into folder ./output/HRNet_relu_aspp/JSTL_large_4/

python test.py

Citation

If you find our work useful or our work gives you any insights, please cite:

@inproceedings{chen2021variational,
  title = {Variational Attention: Propagating Domain-Specific Knowledge for Multi-Domain Learning in Crowd Counting},
  author = {Chen, Binghui and Yan, Zhaoyi and Li, Ke and Li, Pengyu and Wang, Biao and Zuo, Wangmeng and Zhang, Lei}
  booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
  year = {2021}
}

GitHub

https://github.com/Zhaoyi/Yan-DKPNet