Computer vision clean-up (archive)

I’ve noticed that iNat has a very strong bias towards suggesting “Primula pauciflora” for any Dodecatheon/Primula photo posted for this area and the surrounding large area (i.e. Alberta and the contiguous northern states), regardless of whether any identifying characteristics are visible in the observation record, which is often just a single photo of the whole plant taken from some distance away, probably unidentifiable beyond genus/section. This probably contributes to some false distribution data, when people simply select the suggestion, often I would guess without realizing there may be one or more other superficially similar species, and without awareness of the fine details that may need to be examined to distinguish them.

Another example is for Gaillardia, in which the iNat AI has a strong bias towards suggesting Gaillardia pinnatifida, which is way out of range for this area.

As noted previously, it seems the tendency for AI to “try” to be overly precise may be counterproductive in some cases. On the other hand, maybe the error will fix itself with time in the case of the Gaillardia, as more observations are made for this area and corrected as to species?