bigknight / sanet_implementation Goto Github PK

请问一下你的密度图是怎么生成的呢？我的运行结果和你的一样，对于所有输入的image都倾向于输出固定的结果（0）。你的这种情况有没有得到改善呢？后来是怎么解决的呢？

Hello , thanks for your open-source.
I am try to reimplement the result of this paper, and I am not sure about your experiment environment such as Tensorflow version and dependency.
Would you plz give me more detail about this?
And step information about training and testing.
Thank you very much.

1. 论文原文是1×1的卷积，不是2×2的卷积
   

   SANet_implementation/SANet_model.py

   Lines 31 to 32 in a6581b4

   	with tf.variable_scope(name_or_scope="branch_1x1_" + str(layer_number)):
   	branch_1x1 = slim.conv2d(data_input, channel_output, [2, 2], 1, "same")

2. 关于池化
   

   SANet_implementation/SANet_model.py

   Line 71 in a6581b4

   feature_map_encoder = slim.max_pool2d(feature_map_encoder, [2, 2], 2, "SAME", scope="max_pooling_2x2")

   
   原文： 2 × 2 max-pooling layers after each module to halve the spatial resolution of feature maps.
   所以我认为应该是VALID

另外还有一个问题我没弄明白，望解答
3. 此处的2*channel_output与原文的意思是否一样，原文的feature dimensions指的是什么？

您好，训练用的tensorflow哪个版本，谢谢

非常感谢您的开源，我看了一下您在制作密度图标签的时候，采用的是固定核，核的大小取的是5，但是我看其他介绍，他们取的是15，这个对结果有影响吗？

你好，请问论文中Evaluation details中提到测试时对图片的处理，具体怎么实现的呢？谢谢！
ps.我只简单跑了网络结构 效果不是很好 可能训练或者测试细节没加上去的原因 另外发现加上IN层虽然加速收敛了 效果相对而言却差了

原文：For testing the model trained based on patches, we crop each test sample to patches 1/4 size of original image with 50% overlapping. For each overlapping pixels between patches, we only reserve the
density value in the patch of which the center is the nearest to the pixel than others, because the center part of patches has enough contextual information to ensure accurate estimation.