Our cameras capture light at 16bit depth, and single float (32bit) format stores values in 23-bit precision, so I don't think the use of double is justified. Since double precision takes exactly twice as much space as single precision float, and a typical datacube is 256px by 256px by 2048 frames, the 3D mask here would take up 1 GiB of memory instead of just 500MiB. https://github.itap.purdue.edu/Nolte-Group/alpha-prime-system/blob/9d9036ad83037227526c65904efc6b057a1cd39b/onekey4-core/central_mask.m#L22