Yes sounds like a challenge.
Maybe they could implement a much smaller model for this, which is only good in detecting the trade screen (which is a pretty unique screen) and then let the normal model run to find the cards.
Then they would have the info if its a trade or not and which cards.
I dont know how the current model woks, but maybe they also already have the info about who got which card (first card detected and second card detected)
(Multiple pictures means running this workflow multiple times?)
Or you could base it not on the trade screen, but on the trade info table which is available in the app, so it just has to reed the table.
I think the most user friendly would be if this functionality is behind the normal scan button, but I dont think that where this is is really important.