CSRNet-Paddle

This is a simple and clean implemention of CVPR 2018 paper "CSRNet: Dilated Convolutional Neural Networks for Understanding the Highly Congested Scenes".

Requirement

1. PaddlePaddle 1.6.2

2. Python 3.6

Data Setup

1. Download ShanghaiTech Dataset from Baidu(code:y7qt)

Train

1. Download the parameters of the pre-training model VGG16 (vgg16.pkl) from Baidu disk(code:4vh3)
2. Run train.py

Testing

1. Run test.py for calculate MAE of test images .

Other notes

1. We trained the model and got 14.19 MAE and 484.35 MSE at 471-th epoch on ShanghaiTech PartB.
2. For some reasons , the result is still a bit behind the paper. If you get better result , I hope you can tell me.