A simple and unofficial Keras version implementation of Multi-column CNN for crowd counting.

Multi-column CNN is the crowd counting algorithm proposed in a CVPR 2016 paper "Single Image Crowd Counting via Multi Column Convolutional Neural Network".

The data can be downloaded on dropbox or Baidu Disk can't be used directly without some preprocessing.

1. Create directory data/original in the root path of this repository, then move the decompressed ShanghaiTech to it.

2. Run the create_gt_test_set_shtech.py to generate the csv files for test which can be loaded as:

   

3. Run the create_training_set_shtech.py to generate selected images and csv files randomly for training and validation. in formatted_trainval.

These three python files in data_preparation are adapted from the original MATLAB version preprocessing implemented in this mcnn repository in pytorch.

Some outputs after 200 epochs -- loss curves:

• Good one:

• Acceptable one:

• Bad one: