The End of Deconvolutions

Deconvolutions were introduced in 2014 in “Fully Convolutional Networks for Semantic Segmentation” and has been extensively used in Semantic Segmentation and Generative Adversarial Networks. But its saturated now and the problems involved with it including checkerboard effects play a huge role in the error it produces. This blog post goes down the journey of deconvolutions and problems associated with it. It also suggests some solutions and how it can be replaced by better alternatives such as subpixel-cnn. If you are doing segmentation or working with generative networks, its time to move away from deconv.