The layer, run forwards. The kernel — a volume the full depth of the input — slides across the spatial positions, and each stop writes one cell of one feature map. When the gold kernel has visited all nine positions its map is finished, and the second filter takes over: same input, its own weights, drawn in purple — and it writes a new map in front of the first. Two filters, two output channels; that is all C_out means. The maps hold response values, not colours — the colour here only records which filter wrote which map.