On the memory footprint of FSD

[ZeaZoM]

I read your article 'Fully Sparse 3D Object Detection' and it inspired me a lot. I have a small question to ask: in Figure 5 you show the memory footprints and inference latency in different perception ranges, I found in the figure that when the detection distance is relatively close (e.g. 50m and 100m), the memory usage of FSD is higher than that of CenterPoint. I am not sure why, can you provide the rationale behind this?