CSRNet-Keras

An unofficial implementation of CSRNet for crowd counting in Keras-Tensorflow.

Paper:

• Original_paper: CSRNet: Dilated Convolutional Neural Networks for Understanding the Highly Congested Scenes.

Results now:

Weights can be downloaded in the release.

Dataset:

• ShanghaiTech dataset: dropbox or Baidu Disk.

Training Parameters:

1. Loss = MSE;

2. Optimizer = Adam(lr=1e-5);

3. Batch size: 1;

4. Data augmentation: Flip horizontally online randomly;

5. Weights: Got best weights of SHB in epoch 135, the best one of SHA in epoch 127, and here is the loss records:

   • SHA:

   

   • SHB:

   

6. Prediction example:

   

   

Run:

1. Download dataset;
2. Data generation: run thegenerate_datasets.ipynb .
3. Run the main.ipynb to train, test, analyze and evaluate the image quality.