I noticed in the mono-camera deformable attention code, the cam2img transformation uses only camera intrinsics, without extrinsics for BEV-to-camera conversion. Could you explain why this design works?

Thanks for your insight!