Two new species classifiers have been released.
Southern Africa V3.0.0
Those who have been using TrapTagger for a long time will remember that our first species classifier was a Southern-African model which was replaced by the general sub-Saharan model that’s currently available. In an effort to eke out some additional performance, we have made the decision to split our African model into separate regional models again – which should be released over the coming months with an East-African model currently in training.
This model drops non-endemic species like gorrilla, bongo, red river hog, giant forest hog, forest buffalo, Thomson’s gazelle, Grant’s gazelle, kob etc. in favour of individual duiker species (blue, common, red, yellow-backed), black wildebeest, bontebok, hedgehog, meerkat, pangolin, puku, suni, and rock hyrax. On top of the additional species coverage, this model also benefits from more-extensive coverage of domestic species (adding donkey, horse, caprine (sheep/goat), cat, dog and pig) and other common categories like birds, rodents, leporids (hares & rabbits), squirrels, otters and lizards through cross-pollination from our global training data pool.
This Southern-African model roughly covers all regions in Angola, Zambia, Malawi, Mozambique and southward, but those in other regions will likely find that this model will be substantially more accurate than the sub-Saharan model on common species like lion, leopard, zebra, giraffe etc. However, this is obviously at the expense of it’s ability to identify non-endemic antelope etc. As such, it could be used depending on your particular species of interest. As usual, it’s always worth trying it out on some test data before using it in earnest.
(Photo credit: Humboldt Universität zu Berlin)
South America V2.0.0
This classifier has actually been available since December 2024 – so if you have imported a survey since then then you have been using it already. We just held off on announcing it at the time as we didn’t want it to be overshadowed by the archival update (so now it’s being overshadowed by the waterhole update instead).
This model was trained on a significantly more-varied training dataset from a lot of different users covering a wider variety of regions, camera models, camera setups etc. and thus should be a lot more robust and precise. Like the Southern-African model, it also benefits from cross pollination from our global training dataset giving it better coverage of domestic species (donkey, sheep, goat, dog, pig) along with useful broad categories like bird, leporid, otter, crocodilian, primate and squirrel. Moreover, we were able to split out the various agouti (Azara’s, black-rumped, black, central-american, and red-rumped), brocket (pygmy, gray, red, and dwarf) and peccary (white-lipped, collared, and chacoan) species into individual categories.
Most significantly, this new model adds a whole host of new species: guinea pig, white-eared opossum, marsh deer, andean mountain cat, short-eared dog, grey fox, south american tapir, margay, lowland paca, crab-eating fox, culpeo, big-eared opossum, white-tailed deer, pampas fox, common opossum, viscacha, laminid, southern tiger cat, pampas cat, grison, geoffroy’s cat, spectacled bear, four-eyed opossum, crab-eating raccoon, and mountain paca.