Not quite understanding the occupancy along the height

[yaobaishen]

Question about the FlashOCC paper:

After extracting the image 2D feature, most of the subsequent modules, except the final Channel2Height operator, operate on the 2D BEV feature, how does the model reconstruct the occupancy status along the height dimension? For example, in the case of the traffic light crossbar over the road, the entire height of the corresponding BEV grid will be occupied, as it's reconstructed from 2D BEV feature?

I realize this might be a simple question but cannot figure it out, could anyone show insight? thanks in advance!