Computer Vision News - February 2018

The inception network runs its input in parallel through two types of operations: convolutions of different scales and max pooling -- then concatenates their outputs. Computer Vision News We Tried for You: TensorFlow-Slim 13 Tool To illustrate how the inception unit works, we’ll first look at a naïve implementation: in the figure at the right, the input is passed for parallel processing to three convolution units and a max pooling unit. The outputs of all 4 parallel units are concatenated together to form an output matrix of massive depth. To reduce the volume of computations, the actual implementation includes conv units with 1x1 filters before each conv unit with non-1x1 filters and after the max pooling unit, cutting the number of operations in half. Inception-v2 is a variant of Inception-v1. It uses smaller conventional kernels, for example replacing the 5x5 conv unit by two chained 3x3 conv units, which reduces the number of parameters, memory use and number of operations.