Either crop or pad the input volume: What is beneficial for Convolutional Neural Network?

Convolutional Neural Network (CNN) is the most popular method of deep learning in the machine learning field. Training a CNN has always been a demanding task compared to other machine learning paradigms, and this is due to its big space of hyper-parameters such as convolutional kernel size, number of strides, number of layers, pooling window size, etc. What makes the CNN's huge hyper-parameters space optimization harder is that there is no universal robust theory supporting it, and any work flow proposed so far in literature is based on heuristics that are just rules of thumb and only depend on the dataset and problem at hand. In this work, it is empirically illustrated that the performance of a CNN is not linked only with the choice of the right hyper-parameters, but also linked to how some of the CNN operations are implemented. More specifically, the CNN performance is contrasted for two different implementations: cropping and padding the input volume. The results state that padding the input volume achieves higher accuracy and takes less time in training compared with cropping method.