Open camera-trap AI fine-tuned on one GPU outperforms stock models on UK wildlife

A preprint posted Oct. 23 by Paul Fergus reports fine-tuning the open-weight DeepFaune camera-trap classifier on 28 UK wildlife classes using a single GPU, motivated by distinctions regional conservation needs such as native red versus invasive grey squirrels. On 20,212 hand-verified images from the MammalWeb citizen-science project, the resulting DeepFaune-UK v1.2 labelled 88.8% of images exactly right, against 56.5% for DeepFaune v1.5.0 and 59.0% for SpeciesNet, though DeepFaune v1.5.0 scored 91.4% at its own coarser labels, a difference the paper found not significant. Fine-tuning on 10% of the training images (about 150 per class) gave similar accuracy, one such run took about 22 minutes on under 4 GB of GPU memory, and the paper says the largest like-for-like gain came from a newer base model released by others.