On the use of variable stride in convolutional neural networks

This paper explores the idea of changing the stride value in convolutional neural networks depending on the position of the pixel within the image: a smaller stride value is used when processing the center of the image, while a larger one is used for pixels close to the edges. We show several examples of image classification tasks where the proposed approach outperforms a baseline solution of same computational cost using fixed stride and several counterexamples where it does not - and explain why this is so. The proposed method has been successfully tested using several contemporary datasets and can be easily implemented and extended to other image classification tasks.