Junyu Gao

Line 46 47 in https://github.com/gjy3035/PCC-Net/blob/ori_pt1_py3/datasets/shanghaiTechB.py

由于数据集没有明确划分验证集，因此我们在train 的过程中使用test数据作为指标的监控。 严格来讲，是需要划分验证集的。建议从train中划分10%的数据。事实上，我们论文也是按照该方案执行的。为了保证大家的可复现性，才在最终release代码时，才用了test（毕竟每个人随机划分出的验证集是有差别的，这也会影响复现效果）。 test.py则是为了最终的评估和可视化。

```cclabeler/users/test.json```: {"password": "optimal", "data": ["1", "2", "3"], "done": ["1"], "half": ["2"]} means: username: test passed: optimal

We will release our code after 11.17. Thanks for your attention!

不能自动标注，仅支持人工标注

Please use this code：https://github.com/gjy3035/NWPU-Crowd-Sample-Code

Firstly, comparing the attention modules, SCAR is almost the same as DA Net. Then, we need to make explanations for it: Our original paper is inspired by SCA cnn (ICCV2017)...

Other links: http://share.crowdbenchmark.com:2443/home/NWPU-Crowd_Dataset https://pan.baidu.com/s/1c2eLEE7leN0jz-fM38zyIQ#list/path=%2F

Thanks for your attention! The model is uploaded at http://share.crowdbenchmark.com:2443/home/Pre-trained_Models_NWPU-Crowd