# Atrous (Dilated) Convolution: How one can compute responses of arbitrarily high dimensions in DCNN?

According to this paper (page 4, bottom-right), atrous convolutions can be used to compute responses of arbitrarily large dimensions in Deep Convolutional Neural Networks.

I do not understand how something like this is true, since by upsampling the filters, one effectively can apply the filter less times to an image, unless one also upsamples the image. Applying the filter less times as I see it obviously means that the output (response) will be of lower dimensionality.

Is there something that I am missing here?