Deep Learning for Computer Vision

Training a fully connected layer as a convolution layer

The problem with the sliding window is the computational complexity. The complexity is because predictions are made for every window. Deep learning features have been computed for every window for overlapping regions. This computation of features for overlapping regions in cropped windows can be reduced. The solution is to use a fully convolutional net which computes the feature only once. For understanding a fully convolutional net, let's first see how to convert a fully connected layer to a convolution_layer. The kernel is changed to the same size, with the same number of filters as the number of neurons. It can be repeated for other layers too. Changing the kernel size is an easier ...