I’ve just tested out this cool insect detector and segmentation AI tool — FlatBug https://lnkd.in/dHQGugUm 🐞.
It works quite well! But the results may be a bit biased, as I couldn’t find any open-source insect dataset that wasn’t already part of the model’s training and validation splits.
I randomly selected 100 images from the ArTaxOr Kaggle collection (https://lnkd.in/dn6MAeQ5) and ran them through the FlatBug large model (you can also choose the medium and small variants: flat_bug_M.pt and flat_bug_S.pt).
The CLI example (fb_predict -i path/to/images -o path/to/results -w flat_bug_L.pt) auto-downloads the weights and gets you started immediately.
Some notes: these are crisp images with the insect in focus. When insects appear in a bokeh background, the detector seems to struggle, same with string-like objects such as roots, which can look like spider or insect legs.
From my tests, filtering out any predictions below 0.8 confidence would remove almost all false positives (just 1–2 misdetections), though it also misses a fair number of true positives. You’ll see some false positives and even images with no detections in this GIF.
Is it perfect? No, especially since I may have tested on data the model has already seen. But it’s a great start and a handy tool I can recommend if you work with insects. I hope to see it grow and improve!
