Computer Vision News - October 2018

Let’s look at the network and the five component parts we want to take a closer look at. In the code snippets below we have left the line numbers from the original file, so that you can go look at them in context and understand their function in the code. The Encoder The backbone of the model is the implementation of the modified aligned xception network (marked I. DCNN). This implementation consists of three “flows”, as seen in figure 4 of the article, reproduced below: the entry flow comprised of 3 blocks (lines 347-363), the middle flow comprised of 16 repetitions (lines 366-370) and the exit flow comprised of 2 blocks (lines 371- 376). Line 359 is crucial to the implementation -- it saves the output of the second entry layer/block under the variable skip1, this is in fact the heart of the encoder, as it is concatenated to preserve spatial features, as you will see later in the code. Focus on… 57 Tool Computer Vision News