Single Image Crowd Counting via Multi Column Convolutional Neural Network

This is an unofficial implementation of CVPR 2016 paper "Single Image Crowd Counting via Multi Column Convolutional Neural Network"

Installation

  1. Install pytorch
  2. Clone this repository 
    git clone https://github.com/svishwa/crowdcount-mcnn.git

    We'll call the directory that you cloned crowdcount-mcnn ROOT

Data Setup

  1. Download ShanghaiTech Dataset from
    Dropbox: https://www.dropbox.com/s/fipgjqxl7uj8hd5/ShanghaiTech.zip?dl=0

    Baidu Disk: http://pan.baidu.com/s/1nuAYslz

  2. Create Directory 
    mkdir ROOT/data/original/shanghaitech/
  3. Save "part_A_final" under ROOT/data/original/shanghaitech/
  4. Save "part_B_final" under ROOT/data/original/shanghaitech/

  5. cd ROOT/data_preparation/

    run create_gt_test_set_shtech.m in matlab to create ground truth files for test data

  6. cd ROOT/data_preparation/

    run create_training_set_shtech.m in matlab to create training and validataion set along with ground truth files

Test

  1. Follow steps 1,2,3,4 and 5 from Data Setup

  2. Download pre-trained model files:

    [Shanghai Tech A]

    [Shanghai Tech B]

    Save the model files under ROOT/final_models

  3. Run test.py

    a. Set save_output = True to save output density maps

    b. Errors are saved in output directory

Training

  1. Follow steps 1,2,3,4 and 6 from Data Setup
  2. Run train.py

Training with TensorBoard

With the aid of Crayon, we can access the visualisation power of TensorBoard for any deep learning framework.

To use the TensorBoard, install Crayon (https://github.com/torrvision/crayon) and set use_tensorboard = True in ROOT/train.py.

Other notes

  1. During training, the best model is chosen using error on the validation set. (It is not clear how the authors in the original implementation choose the best model).

  2. 10% of the training set is set asised for validation. The validation set is chosen randomly.

  3. The ground truth density maps are obtained using simple gaussian maps unlike the original method described in the paper.

  4. Following are the results on Shanghai Tech A and B dataset:

            |     |  MAE  |   MSE  |
        ------------------------
        | A   |  110  |   169  |
        ------------------------
        | B   |   25  |    44  |

  5. Also, please take a look at our new work on crowd counting using cascaded cnn and high-level prior (https://github.com/svishwa/crowdcount-cascaded-mtl), which has improved results as compared to this work.