1，首先，DispNet去除了corr層，因為我還不確定corr層的原理是否正確，然後繼續採用左右影像cancat在一起，然後現在採用FlyingThings3d的資料集，來看看效果如何？

2，我覺得road detection的任務還是不能用分類的方法，分類的方法還是產生單個單個的結果，很容易產生不連續的道路預測結果。

3，在參考了幾個很重要的文章之後，自己模仿著寫出了新的一個網路，最重要的模組程式碼如下：

def layer(x,filter_out):
  ksize1x1 = 1
  ksize3x3 = 3
  ksize5x5 = 5
  ksize7x7 = 7
  stride = 1
  filter_out_32=32
  filter_out_1=1
  shortcut = x
  with tf.variable_scope('block_A'):
     x1_conv_1x1=conv_2d(x, ksize1x1 ,stride,filter_out)
     x1_conv_3x3=conv_2d(x, ksize3x3 ,stride,filter_out)
     x1_conv_5x5=conv_2d(x, ksize5x5 ,stride,filter_out)
     x1_conv_7x7=conv_2d(x, ksize7x7 ,stride,filter_out)
     x2_conv_1x1=Max_pooling(x1_conv_1x1, pool_size=[3,3], stride=1, padding='SAME') 
     x2_conv_3x3_1x1=conv_2d(x, ksize1x1 ,stride,filter_out_1)
     x2_conv_5x5_1x1=conv_2d(x, ksize1x1 ,stride,filter_out_1)
     x2_conv_7x7_1x1=conv_2d(x, ksize1x1 ,stride,filter_out_1)
     cancat1=tf.cancat()
     cancat1=cancat1+shortcut 
     shortcut =cancat1
     cancat1= Relu6(cancat1)
  with tf.variable_scope('block_B'):
     x1_conv_1x1=conv_2d(cancat1, ksize1x1 ,stride,filter_out)
     x1_conv_3x3=conv_2d(cancat1, ksize3x3 ,stride,filter_out)
     x1_conv_5x5=conv_2d(cancat1, ksize5x5 ,stride,filter_out)
     x1_conv_7x7=conv_2d(cancat1, ksize7x7 ,stride,filter_out)
     x2_conv_1x1=Max_pooling(x1_conv_1x1, pool_size=[3,3], stride=1, padding='SAME') 
     x2_conv_3x3_1x1=conv_2d(x1_conv_3x3, ksize1x1 ,stride,filter_out_1)
     x2_conv_5x5_1x1=conv_2d(x1_conv_5x5, ksize1x1 ,stride,filter_out_1)
     x2_conv_7x7_1x1=conv_2d(x1_conv_7x7, ksize1x1 ,stride,filter_out_1)
     cancat2=tf.cancat()
     cancat2=cancat2+shortcut)
     cancat2= Relu6(cancat2)

  return cancat2