DeepFaune-UK v1.1: ONNX Model Weights for UK Mammal and Bird Species Detection and Classification

Description:

DeepFaune-UK is a two-stage detection and classification pipeline for automated identification of UK wildlife in camera trap imagery, developed as part of the TrapTracker platform (traptracker.co.uk). A YOLOv10x detector locates animals, people, and vehicles in frames; a dedicated classifier then assigns one of 30 species/class labels to detected animal and person crops. Vehicles are filtered at the detection stage and are not passed to the classifier. Weights are exported in ONNX format for cross-platform inference without requiring the original training framework.

Supported classes (30):

Erinaceus europaeus (hedgehog), Sciurus carolinensis (grey squirrel), Sciurus vulgaris (red squirrel), Capreolus capreolus (roe deer), Cervus elaphus (red deer), Vulpes vulpes (fox), Meles meles (badger), Person, Capra hircus (goat), Oryctolagus cuniculus (rabbit), Columba palumbus (wood pigeon), Phasianus colchicus (pheasant), Ovis aries (sheep), Passer domesticus (house sparrow), Dama dama (fallow deer), Bos taurus (cattle), Felis catus (cat), Martes martes (pine marten), Canis familiaris (dog), Calibration pole, Accipiter gentilis (goshawk), Buteo buteo (buzzard), Capercaillie (cock), Capercaillie (hen), Numenius arquata (curlew), Numenius arquata chick, Equus caballus (horse), Corvus corone (carrion crow), Sus scrofa (wild boar), Muntiacus reevesi (muntjac).

Training data:

The model was fine-tuned on 38,413 labelled animal/person crops drawn from UK camera trap images collected through conservation deployments over nearly a decade, contributed by multiple partner organisations. The dataset spans daylight RGB and infrared night-time images, motion blur, partial occlusion, multi-animal frames, juveniles and adults, manufacturer overlays/timestamps, and varied UK habitats and deployment settings.

Raw training images are not released, as many were provided under partner-specific data agreements that pre-date this open model release.

Training setup:

Architecture: Vision Transformer ViT-L (DINOv2), fine-tuned from the DeepFaune European camera-trap backbone; Approach: transfer learning — linear-probe stage followed by full fine-tuning of a fresh 30-class UK classification head; Input resolution: 518 × 518; Training hardware: NVIDIA RTX A6000; Framework: PyTorch / timm; Export format: ONNX

Performance:

On a held-out UK test set of 5,766 crops, DeepFaune-UK achieves a macro-averaged F1 score of 0.9945 (overall accuracy 0.994) across all 30 classes. Performance is strong and even across species; the lowest-scoring classes are closely related deer — fallow (F1 0.974), red (F1 0.978), and roe (F1 0.987) — reflecting genuine visual similarity rather than systematic error. All other classes score F1 ≥ 0.985.

On out-of-distribution evaluation (camera deployments not represented in training), the pipeline produced high-confidence, ecologically plausible predictions across a 262-image test set. Formal accuracy on OOD data should be validated against ground truth for any specific deployment.

Intended use:

For conservation practitioners, ecologists, and citizen science projects to automate species identification in camera trap datasets. Released freely for non-commercial conservation, research, teaching, and citizen science use, particularly for under-resourced organisations lacking compute or ML expertise to train from scratch.

Acknowledgements:

Fine-tuned from the DeepFaune European camera-trap backbone. Full credit to the DeepFaune initiative — Rigoudy, Dussert, Miele, Chamaillé-Jammes and colleagues — for the foundational work this builds on.

Related publication:

Full methodology, evaluation, and benchmarking described in: arXiv:2606.10940 (link in Related Works as "is supplement to").

License: CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International)

Citation: Please cite arXiv:2606.10940 when using this model in published work. Commercial use requires separate permission from the author.