In these convolutions, the filter is applied over an area larger than its length by skipping input values with a certain step. They are particularly useful in segmentation tasks as they can merge spatial information over larger input areas without losing resolution.