Click a unit in the top feature map (right) and its receptive field appears on the image (left). The two panels share one spatial scale, so each cell of the map covers exactly the pixels it summarises. The table carries the two running quantities: the field adds (k−1)·jump per layer, the jump multiplies by every stride. Click a layer tag to remove it. Neighbouring units' fields sit one jump apart — click along a row and watch the box move by exactly that many pixels. Responses fills the map, computed live: every conv applies one fixed vertical-edge kernel then a ReLU, every pool takes its window maxima — a stand-in for a single learned filter, so read the resolution, not the values.