ByteTrack with 50k detections

[tonyreina]

I know this sounds extreme, but I am creating a device for detecting cells in microscope images and there are ~ 50k cells in one image that I would like to track over the course of 30 frames.

I've managed to get Ultralytics/SAHI and SV to work together for this, but have found that the Detections object grows to around 160 GB in RAM. I have 250 GB of RAM so I am ok for now, but am wondering if there are any out-of-memory tricks for tracking this many detections. For example, if I had a huge dataframe then I could use something like Dask to handle arrays that are too large to fit into memory. Is there something like this for SV?

Thanks.
-Tony