leeyeehoo / csrnet Goto Github PK

Hi, first of all, thank you very much for your work.
But, you didn't elaborate on how to generate a density map in your paper. How can I get a density map from estdmap? How to post-process estdmap?

How to annotate the car dataset and make the corresponding density maps?

您好，非常感谢您能够开源keras代码，方便我们的学习。最近，在复现您的代码的时候，发现在训练过程中 loss 并没有减低，一直维持在一个0.001 左右，非常希望能够得到您的指点。

Hello!
I transfer the Caffe Model to the MXNet implementation.
And I set the mean value [103.939, 116.779, 123.68] in BGR order.
In the inference, divide the image into 4 patches and predict them, then sum the four density maps.

However, I get the result: MAE 74.712, MSE 125.011 in ShanghaiTech Part_A dataset.
It is worse than the result paper reported (MAE 68.2, MSE 115.0).

When feeding the whole image into the model, MAE: 72.585, MSE: 119.049 (mean value: [110.474, 118.574, 123.955]), MAE: 72.626, MSE: 119.541 (mean value: [103.939, 116.779, 123.68]). It is still worse than that paper reported.

And I test the model on Caffe, here is the evaluation code.
Mean Value: [110.474, 118.574, 123.955]
MAE: 72.189, MSE: 118.791 (Caffe CPU)
MAE: 72.189 MSE: 118.791 (Caffe GPU)

Mean Value: [103.939, 116.779, 123.68]
MAE:72.203, MSE: 119.219 (Caffe CPU)
MAE: 72.203 MSE: 119.219 (Caffe GPU)

Could you please provide the mean value and the prediction result?

Thank you!

首先。图像预处理是指将所有待训练和测试的图片根据标注的人头点的坐标生成对应的密度图，并与标注的人头总数一起作为ground-truth。训练阶段是指将所有的训练集（包括图像预处理生成的ground-truth）输送到以VGG16前十层作为前端的网络进行人头特征提取，然后将提取到的人头特征输送到空洞卷积神经网络的进行训练，最后通过提取出的人头位置特征生成对应的密度图。
麻烦看一下我理解的对吗？谢谢

你好，请问caffe版本输入图片的transform和pytorch版本是一致的么？为何我在caffe版本使用和pytorch相同的transform得出的结果完全不同？如果不是请问caffe输入需要如何处理？谢谢