Computer Vision News - February 2018

The last five years have seen an explosion in deep learning , which allowed huge advances in almost every application field. For us, one of its main benefits is its effect on various visual tasks. Together with these developments, the explosion in ready-to-use libraries for implementing deep learning might be confusing when contemplating which library will best suit your particular needs. This month we are having a look at TensorFlow-Slim , focusing on the VGG and GoogLeNet inception network implementation. First, let’s briefly recap GoogLeNet and inception units . The idea of GoogLeNet is constructing a network from basic convolutional units. This should make the network computationally more efficient. The network is constructed by serially chaining together small, efficient units known as inception units, with only a single fully connected layer at the end. Auxiliary classifier units are added along the chain, to allow viewing intermediate results for improved training of the network based on these results. GoogLeNet has 12 times fewer parameters compared to AlexNet , despite being a much deeper network. Google has published 4 versions of GoogLeNet over the past two years, with the main difference between versions being variations in the inception units they use. Let’s look at the basic structure of inception units: Inception-v1: The idea behind the inception unit is that it is a miniature network, so you can actually think of GoogLeNet as a network made up of a multitude of tiny networks. 12 Computer Vision News We Tried for You: TensorFlow-Slim Tool by Assaf Spanier “TF-Slim wrappers reduce a very lengthy, complex code in TensorFlow to just a few lines”