Does Computer Vision take into account all stages of a holometabolous insect?

General
1

Just wondering, because the same species can look very different at different stages of metamorphosis. But i guess this question could also apply to hemimetabolous insects

2

I know it can identify eg both larval and adult butterflies. Probably it is “aware” that there are multiple ways a species can appear.

1 Like
3

Yes, but outside of Lepidoptera, there are likely not many non adult photos submitted, meaning a high risk of misidentification by the computer vision.

Even within leps, it’s going to be limited by the more observed taxa.

4 Likes
4

Yeah, of course it will be limited by availability of photos. But in theory, it can and does recognize that a taxon can have multiple looks.
I’d argue that it’s similar to recognizing eg an acorn and an oak leaf as both Quercus, which it definitely does.

2 Likes
5

This is a problem in hover flies. The larvae are quite distinct to subfamily, and can apparently be ID’d to genus in at least some cases. But since the CV model isn’t trained on observations of parent taxa when at least one child taxon qualifies for inclusion (as many hover fly species do), almost no larval photos make it into the model. Accordingly, photos of syrphine larvae are almost always suggested to be caterpillars. Increasingly, I think the AI should be trained on at least some “redundant” parent taxa to improve the quality of suggestions.

3 Likes
6

The Large Sand Scarab CV seems pretty good at adults and larvae, as both have pretty large example sets, even the marks it leaves in the sand when its walking about could possibly be IDed

https://www.inaturalist.org/taxa/408947-Pericoptus-truncatus/browse_photos

but of course CV sees what we show it.

2 Likes
7

I think you’re overstating how it works. In practice the computer vision algorithm may do a good job recognizing the different stages of a taxon, but in theory the computer vision model knows nothing about stages. In particular, annotation data is not incorporated into the model.

Since photos are selected at random, the relative number of photos of each stage is what matters. If the relative number of photos per stage is skewed, the algorithm will not be able to recognize the different stages equally well.

3 Likes
8

Thank you. This was my concern that the model could be skewed in this way

10

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.