NanoDetector

A YOLOv5m animal detector and 25-class species classifier for North American camera-trap imagery, built to run on a Jetson Nano in a battery-powered field enclosure.

NanoDetector is one half of a system for monitoring wildlife where there is no power and no network. The other half is a device: an Arduino-scheduled, satellite-connected trail-camera unit that wakes on a timer, classifies imagery on-device, and transmits the conclusions over Iridium instead of the photographs. A season of imagery collapses from tens of gigabytes to a stream of small structured messages.

🔧 Code, device, docs	github.com/sheneman/NanoDetector
🌐 Project site	nanodetector.ai
🦌 Parent project	WHISL — Wildlife & Humans in Shared Landscapes
💵 Funding	NSF Award #2317537

---

Quick start

NanoDetector is a standard YOLOv5 checkpoint trained on stock, unmodified Ultralytics YOLOv5 — no fork or patch is needed.

import torch
from huggingface_hub import hf_hub_download

weights = hf_hub_download("sheneman/NanoDetector", "NanoDetector_YOLOv5m_v1.pt")
model = torch.hub.load("ultralytics/yolov5", "custom", path=weights)
model.conf = 0.75   # threshold used in field deployment

results = model("camera_trap_image.jpg")
results.print()
print(results.pandas().xyxy[0])   # xmin ymin xmax ymax confidence class name

Or with the YOLOv5 CLI:

git clone https://github.com/ultralytics/yolov5.git && cd yolov5
pip install -r requirements.txt
python detect.py --weights NanoDetector_YOLOv5m_v1.pt --source ./images --conf-thres 0.75

Field deployments use --conf-thres 0.75, far above YOLOv5's 0.25 default: species labels feed a satellite summary, so precision matters more than recall. MegaDetector, not this model, is responsible for not missing animals.

---

Model details

File	NanoDetector_YOLOv5m_v1.pt (169 MB)
sha256	397cd7d62fb38116cea7575bed3a7270b6c8301e434834f792e0bb7eb4081d4c
Architecture	YOLOv5m — 290 layers, 20,949,918 parameters
Input	RGB, 640 × 640
Classes	25
Framework	PyTorch / Ultralytics YOLOv5 @ c23a441
Version	v1 — best.pt snapshot at epoch 438 of a 500-epoch run
Developed by	Luke Sheneman, Research Computing and Data Services, University of Idaho
License	MIT

Training data

1,391,742 labelled images merged from five public camera-trap corpora distributed by LILA BC — IDFG, IDFG2, North American Camera Trap Images, WCS, and Caltech Camera Traps — harmonised into a single 25-class taxonomy (mapping table).

Split	Images
train	1,177,647
val	67,449
test	146,646

The corpus is heavily imbalanced — cattle, elk and deer dominate, while badger has 29 instances in the entire test split.

Training: 500 epochs at 640×640, batch 200, SGD, hyp.scratch-low, 2 GPUs via DDP on the University of Idaho RCDS cluster. Full recipe, exact command, and archived opt.yaml/hyp.yaml: docs/training.md.

Classes

idx	class	idx	class	idx	class
0	human	9	lagomorph	18	badger
1	coyote	10	bobcat	19	unknown cervid ⚠️
2	fox	11	pronghorn ⚠️	20	rabbit
3	mountain lion	12	bighorn sheep ⚠️	21	horse
4	elk	13	unknown ungulate ⚠️	22	unknown cat
5	deer	14	vehicle	23	turkey
6	unknown canid ⚠️	15	wolf	24	domestic sheep
7	domestic dog	16	moose
8	bear	17	cattle

⚠️ unvalidated — no instances (or a single instance) in the held-out splits. Predictions for these classes are unsupported by evidence.

Evaluation

Stock YOLOv5 val.py at 640×640 against this checkpoint.

Split	Images	Labels	P	R	mAP@0.5	mAP@0.5:0.95
val	67,449	89,912	0.947	0.930	0.968	0.902
test	146,646	193,326	0.849	0.835	0.872	0.809

Read the test number carefully. YOLOv5's all row is an unweighted mean over classes. The test split contains exactly one instance each of unknown canid and unknown cervid; both score 0, and two zeros in a 22-class mean cost ~9 points of mAP by themselves. Excluding those two degenerate classes, test mAP@0.5 is 0.959 — consistent with validation. The model does not degrade on test.

Selected per-class test results (full tables: docs/evaluation.md):

Class	Labels	P	R	mAP@0.5
deer	29,495	0.968	0.963	0.989
elk	45,625	0.941	0.963	0.987
cattle	85,800	0.916	0.949	0.976
mountain lion	2,465	0.950	0.965	0.989
bear	3,024	0.943	0.935	0.982
wolf	631	0.941	0.946	0.979
coyote	4,175	0.936	0.950	0.962
moose	3,617	0.936	0.770	0.870
turkey	9,025	0.807	0.925	0.948
human	199	0.908	0.693	0.869

Speed: 4.0 ms/image inference, batch 32, datacenter GPU. Jetson Nano is substantially slower.

Intended use

In scope

• Species-level classification of North American camera-trap imagery
• Edge/offline inference where uploading imagery is impractical
• Second-stage classifier behind a class-agnostic detector such as MegaDetector
• Triage and retrieval across large existing camera-trap archives

Out of scope

• Person detection — see limitations below
• Regions outside western North America
• Species not among the 25 classes (they will be forced into the nearest class, not rejected)
• Individual identification, age, sex, health, or behaviour
• Legal, enforcement, or regulatory determinations without human review

Limitations

1. Human recall is poor (0.69 test / 0.72 val). The model misses roughly three people in ten. Only 199 human instances exist across the 193,326-label test split — humans are ~0.1% of a corpus drawn from wildlife surveys. Do not use this model to detect people, and do not use it as a privacy filter. The deployed device relies on MegaDetector for the person class precisely for this reason.
2. Moose are undercounted — recall 0.77 at precision 0.94, consistently across both splits.
3. Four to five classes are unvalidated (⚠️ above): no held-out evidence exists for them.
4. Turkey over-predicts on test data (precision 0.807 vs 0.954 on validation).
5. Sparse-class metrics are fragile — badger rests on 29 test instances.
6. Geographic scope — trained on western North American imagery; assume poor performance elsewhere.
7. Split independence is unverified. Camera-trap imagery is highly autocorrelated (bursts, fixed backgrounds). If splits were assigned per-image rather than per-camera-station, these metrics overstate accuracy at a genuinely new site. The split procedure is not recorded in surviving artifacts.
8. No open-set rejection — every detection is assigned one of 25 classes; novel species are misclassified rather than flagged.
9. Night/IR imagery was not separately evaluated, though it is common in camera-trap data.

Ethical considerations

Human imagery. Camera traps photograph people. This model has a human class, but its human recall is low, so it must not be used to find and redact people — roughly three in ten will get through. Use a purpose-built person detector, and handle imagery under an appropriate human-subjects and retention policy.

Sensitive species. Near-real-time, georeferenced locations for wolves, mountain lions and bear can enable poaching or inflame conflict. Treat detection streams with the same care as any sensitive wildlife location data; aggregate or delay before public release.

Management decisions. Errors concentrate in rare classes, which are often the classes of greatest conservation concern. Model counts should inform human judgement, not replace it. Decisions with legal or livelihood consequences need human verification of the underlying imagery.

How it's used in the field

The device runs two models in sequence. MegaDetector v5 finds that something is there (animal/person/vehicle) with high recall; NanoDetector says what it is. Outputs are fused by IoU so MegaDetector stays canonical for localisation and for person/vehicle calls while NanoDetector supplies species identity. Detections within a 30-second burst are then normalised by majority vote — never across the human or vehicle classes.

Architecture, firmware, and the fusion pipeline: github.com/sheneman/NanoDetector.

Citation

@software{sheneman_nanodetector,
  author  = {Sheneman, Luke},
  title   = {NanoDetector: An edge AI device and species classifier
             for satellite-connected wildlife monitoring},
  url     = {https://github.com/sheneman/NanoDetector},
  note    = {Model weights: https://huggingface.co/sheneman/NanoDetector},
  institution = {University of Idaho}
}

Acknowledgments

Supported by the National Science Foundation under Award #2317537 and by Research Computing and Data Services at the University of Idaho. Built on Ultralytics YOLOv5 and MegaDetector. Training imagery from LILA BC.