Computer Vision News - October 2018

58 Tool Computer Vision News Focus on… Figure 4 of the article, representing the implementation code above. Another important part of the encoder is the atrous spatial pyramid pooling (ASPP, marked II in the Encoder-Decoder figure above). Here it may be best to start at the end of the code snippet, in line 462 you see the concatenation of variables b0 to b4, where b1, b2 and b3 are the result of the atrous convolution at various rates, as can be seen in lines 452-458, b0 is the 1x1 convolution as can be seen in line 445, and b4 is the updated maxpooling (line 437).