Unofficial implementation of W-Net for crowd counting in Keras.

• Original_paper: Varun Kannadi Valloli, Kinal Mehta. W-Net: Reinforced U-Net for Density Map Estimation.

On dataset ShanghaiTech B

Under development...

• ShanghaiTech dataset: dropbox or Baidu Disk.

1. Loss = MSE * 1000 + BCE * 10;

2. Optimizer = Adam(lr=1e-4, decay=5e-3);

3. Batch size: 1;

4. Data augmentation: Flip horizontally online randomly;

5. Patch: No patch;

6. Batch normalization: No BN layers at present;

7. Weights: Got best weights in epoch248(250 epochs in total), and here is the loss records:

   

8. Prediction example:

   

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