The general domain results in 98% mAP, while the compact domain is 75% (for my data set). I am using a 4K camera to detect objects in real time and would like to be able to download and work with the general model.

Hello PM,

I'm working on an urgent case using custom vision to do the selection of defected solar cell photos from a huge amount of library. Due to security requirement, the customer wants to export the model to on-prem to do the prediction. According to the documentation, it could be slightly less accurate than doing in the cloud. I'm wondering if you have the further explanation regarding this difference and how to improve the precision on-prem. Another question is how will it be charged on-prem prediction if the server can't connect to Azure. Hope to hear the feedback soon.

The offline prediction is very vital because users don't need to connect with cloud for every prediction while they are using just their own modules

It would be very helpful to be able to download the compact model as a managed .DLL that has a single .Classify method. This way we could integrate the real time model into any application.

Maybe add GPU option and/or multicore.