As more people upload and identify leaf-mining insects, plant gallers, etc., more of these species are getting incorporated into the Computer Vision models—and these are all great things of course, IMO.
However, as that happens, Computer Vision begins making only arthropod suggestions in some cases for pictures of plants that have often-reported arthropod associates.
For example, here are screenshots I took this month of the CV suggestions for perfectly fine and ordinary pictures of Trifolium repens, Arctium sp., and Asclepias syriaca (not showing any leaf damage, visible insects, or the like). In all three cases, the CV is suggesting only arthropods, no plant species.
Many, but not all, of these CV arthropod suggestions are photographed in association with the plants in question (so one can see where the CV suggestions are coming from), but none of them were present in the example observations, and I find it especially odd that the CV isn’t suggesting plants at all in the mix of options.
On one hand, there’s an obvious temporary solution for us identifiers, which is to periodically go through arthropod taxa of interest and knock out all the plant observations with another ID (which I do). But that doesn’t seem to be improving the underlying issue with the CV suggestions.
So the reason I’m posting is to draw attention to the issue for the developers and researchers working on the CV models and training. Are these errors always going to be with us? Will things get better over the years if we simply keep making correct identifications on top of all the bad choices people make when misled by CV?
Or are there ways CV can be improved in the short term to reduce these kinds of errors in the model results?
