PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Paul Fergus

arXiv:2610.32573v1cs.CV

TL;DR

Regional conservation needs species distinctions that global camera-trap models may not provide, while building such models at scale is beyond many small teams. This study fine-tunes open DeepFaune classifiers for UK wildlife and finds strong strict-accuracy results against the baselines, while shared-label performance and the conditions for democratised development qualify the gains.

  • Problem

    Global camera-trap models may omit locally important species or merge them into coarse labels, while building models at scale exceeds the resources of many groups.

    Source: §1 · pp. 2–3

  • Method

    The study fine-tunes open DeepFaune classifiers on UK classes, pairs them with unmodified MegaDetector v6, and evaluates them against DeepFaune v1.5.0 and SpeciesNet on verified UK images.

    Source: §3.3 · p. 9, §3.6 · p. 12

  • Results

    DeepFaune-UK v1.2 reached 88.8% strict image-level accuracy versus 56.5% and 59.0% for the baselines, but DeepFaune v1.5.0 led it by 2.6 points at shared labels.

    Source: §5 · p. 22

  • Takeaways & Limitations

    Regional fine-tuning can provide locally useful species vocabulary, while access to this form of AI development depends on continued open-weight releases and their licences.

    Source: §5 · p. 22

    The out-of-distribution benchmark comes from one citizen-science project and is dominated by three species, leaving estimates for two species represented by only three images each unreliable.

    Source: §5.1 · p. 24

Abstract

from arXiv · show

Camera-trap species classification is dominated by large models from well-resourced organisations, which often lack the species-level distinctions regional conservation needs, such as native red versus invasive grey squirrels in the UK. We ask whether open weights let a small team with a modest dataset and a single GPU take a real part in AI development. We fine-tuned the open DeepFaune classifier on 28 UK classes, paired it with the unmodified MegaDetector v6, and compared the resulting DeepFaune-UK v1.2 with DeepFaune v1.5.0 and SpeciesNet on 20,212 hand-verified MammalWeb citizen-science images, after removing sequences sharing any of 1,039 images with our training archive. DeepFaune-UK achieved a strict image-level accuracy of 88.8%, including empty frames, against 56.5% and 59.0% for the baselines, but at DeepFaune's coarser label resolution DeepFaune v1.5.0 was more accurate, though not significantly under contributor resampling (91.4% vs. 88.8%). As DeepFaune v1.5.0 uses a newer DINOv3-based classifier, we applied the identical recipe, without new data, to that classifier. The resulting DeepFaune-UK v1.3 reached 90.6% strict accuracy, narrowed the like-for-like gap to 0.8 points and reduced roe deer misclassified as fallow deer by 69%. On 1,850 images from the 26 contributors with no overlapping images, the three DeepFaune-based models did not differ significantly (90.5%, 90.6%, 89.9% like-for-like). Fine-tuning on 10% of the training crops (about 150 per class) gave out-of-distribution accuracy similar to using all of them, and a DINOv2 run on this subset took about 22 minutes with under 4 GB of GPU memory. Open weights thus make regional model development feasible for small teams, but the largest like-for-like gain came from a newer base model trained by others, so this democratisation depends on large actors continuing to release open weights, under licences that may change.

1 Introduction

Camera traps can support monitoring at scale, but their image volumes create a manual-review bottleneck and regional conservation often requires species distinctions that global models do not provide. This study tests whether a small team can adapt open-weight models to UK wildlife with modest data and compute.

1 Introduction

  • A camera-trap review bottleneck makes image analysis labour-intensive, delays conservation decisions and limits scaling to large or long-running deployments.The paper situates this need against ongoing declines in British terrestrial mammals and ground-nesting birds.Source: §1 · pp. 1–2
  • UK management can depend on distinguishing native red squirrels from invasive grey squirrels, because interventions aimed at one must avoid the other.The passage describes species-selective feeders as one potential application for reliable visual classification.Source: §1 · p. 2
  • Species-level monitoring also matters for deer culling, predator management and other conservation decisions, while large models may return coarse labels or omit locally important species.DeepFaune merges squirrels into one class and lacks capercaillie and curlew classes; SpeciesNet may fall back to broader taxonomic labels.Source: §1 · pp. 2–3
  • The study asks whether open weights let a small team adapt models to regional needs with a modest dataset and a single GPU, treating AI democratisation as participation in development.The paper distinguishes this from democratisation of AI use, profits or governance.Source: §1 · p. 3
  • DeepFaune-UK v1.2 combines unmodified MegaDetector v6 with a DeepFaune classifier fine-tuned on 28 UK classes using 42,719 crops, and v1.3 applies the same recipe to DeepFaune v1.5.0’s DINOv3 classifier.The models are compared with DeepFaune v1.5.0 and SpeciesNet on hand-verified MammalWeb images after removing sequences with images also found in the training archive.Source: §1 · p. 3
  • The study evaluates out-of-distribution performance at species and shared-label resolutions, and measures how data quantity relates to accuracy and fine-tuning resource requirements.Its contributions also include sequence- or contributor-resampled uncertainty estimates and a discussion of the conditions on AI democratisation.Source: §1 · pp. 3–4

2 Related Work

Prior camera-trap research highlights both the challenge of generalising to new locations and the practical value of staged detection and classification pipelines. This section reviews open foundation models, data requirements, alternative recognition approaches, and what open weights do—and do not—change about participation in AI development.

2 Related Work

  • Camera-trap classifiers can overstate performance on held-out images from training locations because they learn repeated camera backgrounds, motivating evaluation at genuinely new locations.Previous platform evaluation also found generally low to moderate species-level recall and recommended users assess models on their own data.Source: §2 · p. 4
  • MegaDetector established a taxon-agnostic first stage that localises animals and discards empty frames, separating broadly transferable detection from region-specific species identification.SpeciesNet combines MegaDetector with a global classifier and geographic filtering, returning higher taxonomic ranks when uncertain.Source: §2 · p. 4
  • Open DINOv2 and DINOv3 weights support downstream fine-tuning, and DeepFaune applies this approach through detection followed by vision-transformer classification of European species.DeepFaune v1.5.0 uses a DINOv3 backbone; the cited licences govern how its weights and related models may be shared and deployed.Source: §2 · pp. 4–5
  • Earlier convolutional-network studies linked low training-image counts to weaker or more variable recall, leaving open whether foundation backbones change regional data needs on independent sites.The cited studies also report reduced accuracy at unseen locations and approaches to estimating image requirements or reducing manual labels.Source: §2 · p. 5
  • Vision-language models offer a route to text-named labels, but camera-trap evaluations found zero-shot biological recognition fragile and site-specific adaptation necessary.The passage presents supervised fine-tuning of an open classifier as the more reliable route evaluated here.Source: §2 · p. 5
  • Open weights may widen participation in adapting models without changing the concentration of large-scale model production, and openness depends on more than downloadable weights.The literature distinguishes AI development from use, profits and governance, and describes openness as composite and graded.Source: §2 · pp. 5–6
  • Conservation infrastructure already uses real-time and batch camera-trap analysis, and Trap Tracker can run user-supplied ONNX models, including DeepFaune-UK v1.2 and v1.3.The cited examples include monitoring at 11 curlew nesting sites in Wales and image triage for the National Hedgehog Monitoring Programme.Source: §2 · p. 6

3 Materials and Methods

The study uses a leakage-controlled training archive and a separate, hand-reviewed MammalWeb evaluation set to compare regional DeepFaune-UK models with two baselines. It also measures how training-set size affects out-of-distribution performance and compute requirements on a single GPU.

3.1 Master Dataset

  • 45,931 images across 28 animal classes formed the master dataset, with people, vehicles and calibration poles excluded from the classifier vocabulary.The 28 classes include 18 mammal species and 8 bird species, with capercaillie and curlew each split into two monitoring-relevant classes.Source: §3.1 · p. 6

3.2 Leakage-Controlled Data Splitting

  • Images were grouped using capture metadata and visual similarity, then allocated by group to training, validation and test sets at a target 80/10/10 ratio.The audit found no held-out image meeting the specified near-duplicate thresholds against training images; the frozen split contains 36,657 training, 4,626 validation and 4,648 test images.Source: §3.2 · p. 8

3.3 DeepFaune-UK v1.2 and v1.3

  • DeepFaune-UK pairs unchanged MegaDetector v6 localization with detector-matched training crops and a two-phase, partially unfrozen DeepFaune classifier.v1.2 starts from DINOv2, while v1.3 changes only to the DINOv3 checkpoint; the detector, crops, splits and training recipe remain identical. MegaDetector v6 was selected over YOLOv8s after recovering approximately 97% versus 84% of annotated animals on the OOD set.Source: §3.3 · p. 9

3.4 Baseline Models

  • DeepFaune v1.5.0 and SpeciesNet were run with their specified release configurations, and models were evaluated at fixed operating points without OOD threshold tuning.DeepFaune v1.5.0 used the example-script threshold of 0.5; SpeciesNet used its default ensemble with UK geographic filtering.Source: §3.4 · p. 10

3.5 Out-of-Distribution Dataset

  • The hand-reviewed MammalWeb evaluation set contains 20,212 images after removing every sequence containing an image that exactly overlapped the master archive.The site-disjoint subset comprises 1,850 images from 26 contributors with no overlapping images; only byte-identical overlaps were checked, so re-encoded or resized copies may remain.Source: §3.5 · pp. 10–11

3.6 Evaluation Protocol

  • Image-level scoring compares exact-species predictions with a like-for-like measure that credits labels at the coarser resolution allowed by DeepFaune’s vocabulary.The image label is the most confident animal detection, a detector label for people or vehicles, or “empty”; empty images count toward accuracy.Source: §3.6 · p. 11
  • Uncertainty estimates resample whole camera-trap sequences, with contributor-level resampling as a stricter check on generalisation across sites.McNemar’s test is reported only as a secondary statistic because its image-independence assumption overstates evidence when images are clustered.Source: §3.6 · p. 12

3.7 Learning Curve and Resource Footprint

  • The learning-curve experiment retrained models on stratified 10%, 25%, 50% and 100% subsets, using three seeds per subset and unchanged validation and test splits.Training time, peak GPU memory and GPU energy were recorded; DINOv3 comparisons repeated the 10% and 100% conditions.Source: §3.7 · p. 13

4 Results

DeepFaune-UK v1.2 substantially outperformed both baselines at strict species-level accuracy, while DeepFaune v1.5.0 led at the coarser like-for-like label resolution. Applying the same fine-tuning recipe to the newer DINOv3 classifier improved DeepFaune-UK v1.3; performance changed little with training-set size, and the site-disjoint subset showed no significant differences among the DeepFaune-based models.

4.1 Out-of-Distribution Comparison

  • 88.8% strict accuracy for DeepFaune-UK v1.2 exceeded SpeciesNet’s 59.0% and DeepFaune v1.5.0’s 56.5% on 20,212 OOD images.At DeepFaune’s coarser label resolution, DeepFaune v1.5.0 scored 91.4% versus 88.8% for v1.2; its advantage was less certain under contributor resampling (p = 0.08). Per-species strengths differed, with v1.2 leading for grey squirrel, hedgehog, woodpigeon and carrion crow, while v1.5.0 led for several others, including roe deer and cat.Source: §4.1 · pp. 13–14, Table 1 · p. 14

4.2 Operating Thresholds

  • 0.889 like-for-like accuracy after threshold calibration was nearly unchanged from the default 0.888, and the difference from DeepFaune v1.5.0 remained.Calibration reduced empty-frame animal labels and improved cat F1, but lowered macro recall; the remaining model difference was significant with sequence clustering but not contributor clustering.Source: §4.2 · p. 17

4.3 Following the Upstream Model: DeepFaune-UK v1.3

  • 90.6% strict and like-for-like accuracy for DeepFaune-UK v1.3 narrowed its like-for-like gap from DeepFaune v1.5.0 to 0.8 percentage points.Roe deer labelled as fallow deer fell from 407 to 126 images, but v1.3 retained the same 848 empty-frame false alarms as v1.2; its residual advantage over DeepFaune v1.5.0 was not significant under contributor resampling.Source: §4.3 · pp. 17–18

4.4 Learning Curve

  • With 10% of training crops, v1.2 achieved 0.838 ± 0.007 like-for-like macro-F1, close to 0.839 ± 0.009 with all crops, while DINOv3 models were more accurate than DINOv2 models on the main evaluation set.The observed variation across training fractions was comparable to seed-to-seed variation, so three seeds showed no detectable trend rather than formal equivalence. The backbone comparison used the identical fine-tuning recipe.Source: §4.4 · p. 18, Table 3 · p. 19, Table 4 · p. 19

4.5 Resource Footprint

  • A 10% DINOv2 training run took 22 minutes and used 3.8 GB peak GPU memory, while a full-data run took 44 ± 8 minutes.The reported GPU energy and memory exclude host-computer costs and the cost of pre-training the backbone, detector and baselines.Source: §4.5 · p. 21

4.6 Site-Disjoint Subset

  • On 1,850 site-disjoint images, DeepFaune-UK v1.2, v1.3 and DeepFaune v1.5.0 scored 90.5%, 90.6% and 89.9%, with no significant pairwise differences.The subset covered 26 of 51 contributors and lacked the two rarest species, so its wide intervals cannot resolve one- or two-point differences.Source: §4.6 · p. 21, Table 5 · p. 21

5 Discussion

Regional fine-tuning added UK-specific labels and performed strongly under strict species-level scoring, while the like-for-like comparison with DeepFaune v1.5.0 remained qualified. The discussion also emphasizes that modest adaptation depends on open pretrained models and that the benchmark and resource findings have important scope limits.

5 Discussion

  • 88.8% strict image-level accuracy for DeepFaune-UK v1.2 (90.6% for v1.3), including empty frames, compared with 56.5% and 59.0% for the baselines; v1.3 narrowed the like-for-like gap to 0.8 points.The v1.2 like-for-like comparison was 2.6 points below DeepFaune v1.5.0; the v1.5.0 checkpoint’s backbone, resolution and possible additional data cannot be separated as causes of v1.3’s improvement.Source: §5 · p. 22
  • Regional fine-tuning supplied species-level UK vocabulary, but roe deer–fallow deer confusion remained a principal weakness and was reduced by more than two-thirds with the DINOv3-based checkpoint.DeepFaune-UK also labelled 848 of 2,857 empty images as animals for v1.2; the discussion attributes these errors to detector false positives.Source: §5 · p. 22
  • A 4,270-crop run, roughly 150 per class, performed within seed-to-seed variation of training on ten times as much data and took 22 minutes with under 4 GB of GPU memory.The reported OOD macro-F1 was about 0.85, compared with about 0.99 on in-distribution test crops.Source: §5 · p. 23
  • This is conditional democratisation: small teams can adapt and release regional models, but remain dependent on large actors’ pretrained models and on release terms that may change.The authors describe large and small actors as complementary, with large initiatives providing broad models and regional teams contributing local vocabulary, testing and evaluation.Source: §5 · p. 24

5.1 Limitations

  • The OOD benchmark comes from one citizen-science project, is imbalanced across species, and tests only 14 of DeepFaune-UK’s 28 classes.Some empty-frame reference labels missed animals, and even the contributor-disjoint subset may not be entirely independent because only exact image copies were detected.Source: §5.1 · p. 24
  • Image-level evaluation, a judged like-for-like mapping and fixed baseline operating points keep comparisons reproducible but do not show any model at its best.The baselines’ thresholds were not tuned, and sequence-level information was not used.Source: §5.1 · p. 25
  • Resource estimates use one workstation GPU, exclude host-computer and pretrained-weight production costs, and rely on a single learning-curve recipe with three seeds.The v1.3 analysis was added after the v1.2 benchmark on the same OOD set, so its results are not fully prospective.Source: §5.1 · p. 25

5.2 Future Work

  • Future work should broaden evaluation across projects, regions, seasons and rare classes, while testing more site-diverse deer images, hard-negative empty images and consumer-GPU costs.The authors also propose testing sequence-level aggregation and live performance through Trap Tracker.Source: §5.2 · p. 25

6 Conclusion

The study finds that a small team can develop a regional wildlife classifier by fine-tuning open weights, achieving strong strict accuracy on out-of-distribution images. Comparable accuracy with a newer base model and limited training data supports this approach, while its feasibility depends on continued access to open weights.

6 Conclusion

  • DeepFaune-UK v1.2 achieved 89% strict image-level accuracy on out-of-distribution images, versus 57–59% for general-purpose models, while remaining a few points behind on shared coarse groups.With the newer DINOv3-based classifier and the same regional training recipe, DeepFaune-UK v1.3 reached 91% strict accuracy and narrowed the like-for-like gap to under one percentage point.Source: §6 · p. 26
  • Comparable out-of-distribution accuracy came from 10% of the training data, with DINOv2 training taking about 22 minutes on one GPU using under 4 GB of memory.The DINOv3 run took under an hour and 4.4 GB, with slightly lower macro-F1.Source: §6 · p. 26
  • Open foundation weights can support meaningful AI development by small, domain-led teams, but this depends on continued releases and the licences attached to them.The authors note that licence terms can change between model releases.Source: §6 · pp. 25–26

Funding

The study received no external funding.

Funding

Institutional Review Board Statement

The study used non-invasive camera-trap imagery and involved no direct handling, capture, or manipulation of live animals.

Institutional Review Board Statement

  • The study involved no direct handling, capture, or manipulation of live animals.Images came from camera traps at contributing sites or from MammalWeb; no additional data collection was undertaken.Source: Institutional Review Board Statement · p. 26

Informed Consent Statement

The study did not involve human subjects.

Informed Consent Statement

Conflicts of Interest

The author reports no financial conflict of interest; the referenced platforms are non-commercial research infrastructure, and the benchmarked baselines are third-party systems.

Conflicts of Interest

A Example Images

The example images illustrate model-label differences and characteristic errors, rather than estimating how frequently those patterns occur. They show both v1.3 corrections and failures shared or introduced by the DeepFaune-UK models.

A Example Images

  • The hand-selected examples illustrate result patterns, not their frequency; animal tiles are cropped for display although models processed full frames.Source: A Example Images · pp. 32–33
  • Both DeepFaune-UK versions name exact species where DeepFaune v1.5.0 is limited to coarser groups and SpeciesNet sometimes backs off to family or generic labels.The difficult examples include partial views, night-time infrared and motion blur; DeepFaune v1.5.0 also names raccoon or chamois in two cases.Source: A Example Images · p. 33
  • Starting from the DINOv3 checkpoint reduced roe deer labelled fallow deer from 407 images in v1.2 to 126 in v1.3.The examples also show v1.2 cat or sheep errors on poorly lit or partial animals being corrected.Source: A Example Images · p. 33
  • v1.3 labelled 203 roe deer images as muntjac, versus 42 for v1.2; all three red squirrel images in one sequence were labelled grey by both versions.Source: A Example Images · p. 33
  • The examples also show visible animals reported as empty below the detector threshold and false detections on empty frames labelled as animals.Both versions labelled 848 empty images as animals; changing the checkpoint did not affect this error frequency.Source: A Example Images · p. 33

B Confidence Intervals and the Original Set

The appendix captions specify the confidence-interval procedures and the scope of the original-set comparison. The accompanying examples depict species-level labels, baseline roll-ups, and characteristic errors without supplying new frequency estimates.

B Confidence Intervals and the Original Set

  • Table B1 reports strict-species OOD estimates with 95% cluster-bootstrap intervals from 2,000 resamples of sequences or contributors.Macro-F1 intervals are asymmetric because muntjac and red squirrel each have three images from one sequence and are absent from 35–37% of resamples; when present, all models score F1 zero at this level.Source: Table B1 · p. 33
  • Figure A1 contrasts exact-species labels from both DeepFaune-UK models with the coarser group labels available to DeepFaune v1.5.0.Source: Figure A1 · p. 34
  • Figures A2 and A3 show difficult images where both DeepFaune-UK models name the exact species but neither baseline returns the correct harmonised group.Source: Figure A2 · p. 35, Figure A3 · p. 36
  • Figures A4 and A5 show corrected v1.2 errors alongside remaining failures, including roe deer labelled muntjac and red squirrel labelled grey by both versions.Source: Figure A4 · p. 37, Figure A5 · p. 38
  • Table B2 reports like-for-like OOD estimates and intervals, with calibrated rows using validation-selected thresholds: v1.2 at 0.50/0.30 and v1.3 at 0.50/0.20.Source: Table B2 · p. 39
  • Table B3 gives point estimates on the original 21,640-image set before removing 440 sequences overlapping the master dataset.Source: Table B3 · p. 39

Related papers

Loading 2610.32573v1…