This is a simple and clean implemention of CVPR 2018 paper "CSRNet: Dilated Convolutional Neural Networks for Understanding the Highly Congested Scenes".
-
PaddlePaddle 1.6.2
-
Python 3.6
- Download ShanghaiTech Dataset from Baidu(code:y7qt)
- Download the parameters of the pre-training model VGG16 (vgg16.pkl) from Baidu disk(code:4vh3)
- Run train.py
- Run test.py for calculate MAE of test images .
- We trained the model and got 14.19 MAE and 484.35 MSE at 471-th epoch on ShanghaiTech PartB.
- For some reasons , the result is still a bit behind the paper. If you get better result , I hope you can tell me.