Thanks for sharing these results @jmwright and welcome to the community. That’s a cool setup you implemented, and also an interesting use-case (classifying insects that are challenging due to small dataset sizes).
You’re spot on about the issue you identified, and it’s something we’re aware of. Rather than repeat the discussion that has already happened on this, I’ll point you to the Future Work items we have on “Off Object Observations” and “Out of Model Movements”. Please see the video discussions that are also linked in these for more context.
The TLDR is that the problem is surprisingly nuanced, but we have some ideas for how to address it, detailed in those articles. We aren’t prioritizing a solution at the moment, given our focus on compositional objects, but it would be interesting if you end up testing a solution and find that it works.