Standing on Open Shoulders: Regional Fine-Tuning of Open-Weight Camera-Trap Models for UK Wildlife with Modest Data and Compute

What it claims

DeepFaune-UK beats DeepFaune v1.5.0 and SpeciesNet on UK MammalWeb images only at UK-specific label resolution. When scored at DeepFaune's coarser labels, DeepFaune v1.5.0 scored higher (91.4% vs 88.8% for DeepFaune-UK v1.2), though the gap was not significant when contributors were resampled. Most of the headline 88.8% vs 56.5% gap therefore comes from vocabulary, such as separating red from grey squirrels, roe from fallow deer, and adding classes like curlew and capercaillie that DeepFaune lacks. The author, Paul Fergus (Liverpool John Moores University), concludes that the largest like-for-like gain among the factors varied came from a newer base model trained by others, not from regional fine-tuning.

How it works

The pipeline has two stages. MegaDetector v6 is used unmodified to localise animals and reject empty frames. The open DeepFaune vision-transformer classifier, which uses a DINOv2 backbone in v1.2, is then fine-tuned on 42,719 training crops across 28 UK classes on a single GPU. DeepFaune-UK v1.3 applies the identical recipe, with no new data, to the DINOv3-based classifier from DeepFaune v1.5.0. This separates the effect of the starting checkpoint from the effect of fine-tuning. The learning-curve runs were repeated over three random seeds and report training time, peak GPU memory and measured GPU energy alongside accuracy.

Evidence

• Evaluation set: 1,039 of 21,640 candidate MammalWeb images (4.8%) also appeared in the training archive. Every sequence containing one of them was removed, leaving 20,212 hand-verified images. Confidence intervals and paired tests resample whole sequences or contributors.
• Backbone effect: v1.3 reached 90.6% strict accuracy and narrowed the like-for-like gap with DeepFaune v1.5.0 to 0.8 points. It cut roe deer misclassified as fallow deer by 69%.
• Site-disjoint subset: 1,850 images came from the 26 contributors with no overlapping images. On this subset the three DeepFaune-based models did not differ significantly, scoring 90.5%, 90.6% and 89.9% like-for-like.
• Data efficiency: training on 10% of the crops (4,270, about 150 per class on average) gave out-of-distribution accuracy similar to full-data training.

How to use it

The page states the hardware bar: a DINOv2 fine-tune on the 10% subset took about 22 minutes with under 4 GB of GPU memory allocated. The licences constrain how derived models can be shared and deployed:

• DeepFaune weights are under Creative Commons Attribution-ShareAlike.
• MegaDetector v6 is under AGPL-3.0.
• The DINOv3 backbone under DeepFaune v1.5.0 and DeepFaune-UK v1.3 carries a custom Meta licence.

The fetched text does not give a download location or licence for the DeepFaune-UK weights, and it gives no run commands. The author's earlier UK single-stage detector was released for non-commercial use through the self-hostable Trap Tracker platform.

Limits

The fetched text stops partway through related work. The results, discussion and limitations sections are absent, so per-class figures, including red vs grey squirrel accuracy, are not available here. The evaluation covers MammalWeb citizen-science images only. Scottish wildcat vs hybrid and American mink are named as needs outside the study's scope. The author states that this route depends on large organisations continuing to release open weights under licence terms that can change.