Download the ShanghaiTech dataset, then you also need to generate density map files via generate_map.py. And put them in ./data, then download perspective map from paper Revisiting Perspective Information for Efficient Crowd Counting.
And put the corresonding files inside ./data. You need to move coresponding files under the guidance of config.py. It is mentioned, that you need upgrade to mat files of perspective maps to v7.3 yourself.
Then h5py can read the mat files correctly.
bash download_models.sh
Pytorch version: 1.0 or 1.1
Install the cuda extension, and sh SHA_test.sh.