PCC Net: Perspective Crowd Counting via Spatial Convolutional Network

This is an official implementation of the paper "PCC net" (PCC Net: Perspective Crowd Counting via Spatial Convolutional Network).

In the paper, the experiments are conducted on the three populuar datasets: Shanghai Tech, UCF_CC_50 and WorldExpo'10. To be specific, Shanghai Tech Part B contains crowd images with the same resolution. For easier data prepareation, we only release the pre-trained model on ShanghaiTech Part B dataset in this repo.

Bracnhes

1. ori_pt0.2_py2: the original version.
2. ori_pt1_py3: the current version.
3. vgg_pt1_py3: vgg-backbone PCC Net (higher performance).

Requirements

• Python 3.x
• Pytorch 1.x
• TensorboardX (pip)
• torchvision (pip)
• easydict (pip)
• pandas (pip)

Data preparation

1. Download the original ShanghaiTech Dataset [Link: Dropbox / BaiduNetdisk]
2. Resize the images and the locations of key points.
3. Generate the density maps by using the code.
4. Generate the segmentation maps.

We also provide the processed Part B dataset for training. [Link]

Training model

1. Run the train_lr.py: python train_lr.py.
2. See the training outputs: Tensorboard --logdir=exp --port=6006.

In the experiments, training and tesing 800 epoches take 21 hours on GTX 1080Ti.

Expermental results

Quantitative results

We show the Tensorboard visualization results as below: The mae and mse are the results on test set. Others are triaining loss.

Visualization results

Visualization results on the test set as below: Column 1: input image; Column 2: density map GT; Column 3: density map prediction; Column 4: segmentation map GT; Column 5: segmentation map prediction.

Citation

If you use the code, please cite the following paper: