Why do Waymo and NuScenes have different us_layer_strides even though they use the same CenterPoint?

[JongRok-Lee]

Hello. First of all, thank you for sharing this impressive framework.

My question is as follows: Waymo's CenterPoint-PointPillars performs downsampling at 1x -> 2x -> 4x and then upsampling to match a 1x resolution with us_layer_strides: [1, 2, 4]. However, nuScenes' CenterPoint-PointPillars downsamples at 2x -> 4x -> 8x and matches it to a 4x resolution with us_layer_strides: [0.5, 1, 2].

There is a significant difference in the final output size going into the head, and I'm curious if there is a specific reason for this configuration.