thenukegun10x/TinyFox-2.0

TinyFox-2: High-Performance Night/IR Red Fox & Wildlife Detector

TinyFox-2 is an ultra-fast, edge-optimized 12-class object detection model specialized for nocturnal infrared (IR) camera-traps and perimeter pest/predator monitoring.

Built on the lightweight YOLO26n architecture (2.38M parameters, 5.2 GFLOPs), TinyFox-2 was specifically engineered to solve the historical failure modes of night wildlife detectors: false fires on empty backgrounds, fox-vs-canid confusion, and loss of small quadruped recall.

---

Key Highlights

• 88.2% Fox Detection Rate on Real IR Deployment Video (vs 45.7% in v3 and 80.5% in v4).
• 0% Canid Confusion at Operating Threshold (0.40): Solves the long-standing bug where raccoons, dingoes, or domestic dogs stole fox detections.
• Zero False Alarms on Verified IR Negatives: 0/225 false fires across night-time camera footage.
• Extreme Edge Efficiency:
  • FP32 PyTorch Checkpoint: 5.39 MB (TinyFox-2.pt)
  • FP16 ONNX: 4.75 MB (TinyFox-2_fp16.onnx)
  • INT8 Calibrated ONNX: 2.60 MB (72.5% compression with 0.68 px mean box delta and 100% class agreement).
• Ready for CPU, GPU & Edge NPU Deployment (ONNX Runtime, OpenVINO, TorchScript, Radxa, Jetson, Raspberry Pi).

---

Generational Leap: TinyFox-2 vs. TinyFox 1.0

TinyFox-2 is a complete architectural and algorithmic overhaul compared to the original TinyFox 1.0. Where TinyFox 1.0 was a proof-of-concept single-class detector prone to background false alarms and species confusion, TinyFox-2 is a robust, production-grade 12-class nocturnal wildlife vision model.

Metric / Dimension	TinyFox 1.0 (Legacy)	TinyFox-2 (Current Release)	Real-World Impact
Architecture	Custom NanoDet-Plus Head	YOLO26n (Anchor-Free, Multi-Scale)	Modern backbone with superior feature extraction and gradient stability
Taxonomy & Classes	1 class (red_fox on HF) / 3-class early prototype (fox/cat/dog)	12 distinct biological classes	Full contextual awareness; distinguishes foxes from cats, dogs, mustelids, and marsupials
Fox Recall on IR Video	~38.0% of frames	88.2% of frames	>2.3× improvement in detection continuity across video sequences
Rival Canid Confusion	Severe (stolen by raccoons/dogs)	0.0% rival fires (@ conf 0.40)	Splitting dingo, raccoondog, and domestic_dog completely eliminates misclassifications
False Alarm Rate	5.8% on clean night frames	0.0% (0 / 225 verified negatives)	True zero false-alarm operation on empty infrared scenes
Median Detection Confidence	0.71	0.89	Higher confidence separation between true targets and background clutter
Dataset & Annotation Quality	2,281 images (pseudo-labeled noise, leaking split)	14,627 images (human-verified VOC/COCO, 0-leak stratified split)	Clean ground truth eliminates conflicting loss gradients
Edge Footprint (INT8)	~2.5 MB (experimental)	2.60 MB calibrated via YOLO-Quantizer	72.5% compression with <0.7 px box shift and 100% class match

Why Early Versions Struggled and How TinyFox-2 Fixed It:

1. The Class Trap: In the initial 1-class Hugging Face release, all non-fox animals were forced into the background loss, causing dogs and cats to trigger false fox alerts. In the early 3-class YOLO prototypes (fox, cat, dog), cat was actually wild Asian Leopard Cat and dog was Raccoon Dog, lacking domestic species and confusing distinct canids. TinyFox-2 establishes 12 honest, biology-grounded categories.
2. Canid Splitting Discovery: Grouping raccoon dogs, dingoes, and pet dogs into a generic canid/dog bucket created a decision boundary too loose to separate from foxes. TinyFox-2 explicitly splits dingo, raccoondog, and domestic_dog, which drove rival canid misclassifications on deployment video from 36.5% down to 0.0%.
3. Stratified 0-Leak Training: Earlier iterations suffered from temporal burst leakage across train/val. TinyFox-2 enforces burst-atomic group stratification and uniform IR filtering, ensuring metrics reflect true real-world generalization.

---

Legal & Multi-Factor Licensing Analysis

Releasing an edge AI model trained from heterogeneous wildlife datasets and modern vision frameworks involves multiple legal and intellectual property factors. TinyFox-2 is released under the GNU Affero General Public License v3.0 (AGPL-3.0) with full upstream attribution.

                    ┌────────────────────────────────────────┐
                    │       TinyFox-2 Release License        │
                    │                AGPL-3.0                │
                    └───────────────────┬────────────────────┘
                                        │
        ┌───────────────────────────────┴───────────────────────────────┐
        ▼                                                               ▼
┌───────────────────────────────┐                       ┌───────────────────────────────┐
│     Model Architecture &      │                       │     Upstream Datasets &       │
│        Base Weights           │                       │       Academic Sources        │
├───────────────────────────────┤                       ├───────────────────────────────┤
│ • Framework: Ultralytics YOLO │                       │ • NTLNP (Beijing Normal Univ) │
│ • Pretrained: yolo26n.pt      │                       │   Academic / Research Open    │
│ • License: AGPL-3.0           │                       │ • UNSW Predators / Dryland    │
│ • Commercial: Requires        │                       │   Academic Non-Commercial    │
│   Enterprise License from     │                       │ • ENA24 / LILA BC             │
│   Ultralytics Inc.            │                       │   CDLA-Permissive-1.0 / CC-BY │
│                               │                       │ • Deploy IR Negatives         │
│                               │                       │   User Contributed            │
└───────────────────────────────┘                       └───────────────────────────────┘

1. Model Architecture & Upstream Framework (AGPL-3.0)

TinyFox-2 is built with the Ultralytics YOLO framework and fine-tuned from yolo26n.pt.

• License: GNU Affero General Public License v3.0 (AGPL-3.0).
• Copyleft Requirement: Open-source distributions of this model or downstream network services based on it must make the corresponding source code available under AGPL-3.0.
• Commercial Use Notice: Commercial users who wish to incorporate TinyFox-2 into closed-source or proprietary hardware/software without AGPL-3.0 obligations must acquire a commercial Enterprise License directly from Ultralytics Inc..

2. Training Data Provenance & Academic Use Considerations

TinyFox-2 was trained on a leak-free, stratified corpus of 14,627 nocturnal IR images compiled from:

• NTLNP Dataset (10,338 images, human boxes): Northeast China Tiger and Leopard National Park camera-trap project (Beijing Normal University / Chinese Academy of Sciences). Released for scientific research and wildlife conservation.
• UNSW Dryland Ecology / Predators Dataset (3,524 images, MegaDetector-mined): University of New South Wales predator monitoring in arid Australia. Provided under academic research / non-commercial research terms.
• ENA24 (194 images, human COCO boxes): Eastern North America camera-trap benchmark hosted on LILA BC under the Community Data License Agreement (CDLA-Permissive-1.0).
• Deployment Camera Harvest (225 images): Animal-free infrared CCTV frames captured on deployment hardware, contributed by the author.

Due to the inclusion of academic camera-trap research sets (NTLNP, UNSW), users deploying TinyFox-2 in commercial operations should ensure compliance with upstream academic and non-commercial research terms or retrain exclusively on permissively licensed/owned data (e.g. ENA24 and private footage).

3. Model Safety & Operational Disclaimer

Limitation of Liability & No Warranty (AGPL-3.0 §§ 15–16): TinyFox-2 is provided "AS IS", without warranty of any kind. While TinyFox-2 demonstrates high empirical detection rates (0.983 mAP@0.50 on fox), no autonomous computer vision system is 100% infallible. Do not rely solely on this software for life-critical, commercial livestock protection, or security-critical perimeter defence.

---

Class Taxonomy

TinyFox-2 recognizes 12 distinct animal classes.

The breakthrough from v3/v4 to v5/TinyFox-2 was splitting the generic canid class into separate biological categories (dingo, raccoondog, domestic_dog). Lumping distinct canids caused the detector to learn overly loose decision boundaries that overlapped with foxes. Separating them completely eliminated fox-vs-canid misclassifications.

Class ID	Class Name	Training Instances	Val Instances	Val mAP@0.50	Val mAP@0.50–0.95
0	fox (target)	1,254	159	0.983	0.795
1	bear	369	72	0.978	0.817
2	bigcat	871	121	0.995	0.892
3	cat	787	145	0.972	0.804
4	dingo	836	115	0.956	0.857
5	domestic_dog	96	29	0.693	0.565
6	lagomorph	1,061	300	0.983	0.850
7	mustelid	1,295	241	0.960	0.747
8	possum	1,842	310	0.983	0.915
9	quoll	590	96	0.949	0.844
10	raccoondog	1,085	159	0.971	0.806
11	ungulate	2,749	423	0.976	0.852
All	Macro Average	12,835	2,194	0.950	0.812

(Note: domestic_dog has a low validation sample count in this split. The non-dog classes average 0.973 mAP@0.50).

---

Real-World Deployment Benchmark

Evaluated on 1,443 continuous nocturnal IR frames from camera-trap footage under realistic working distance and low-light noise:

Metric	TinyFox 1.0 (NanoDet)	v3 (dog lumped)	v4 (canid lumped)	TinyFox-2 (v5 Split)
Fox Top Prediction Rate (@0.40)	~38.0%	45.7%	79.9%	88.2%
Rival Top Prediction Rate	N/A (1-class)	36.5%	4.6%	0.0%
Same-Object Dual-Class Fire	N/A	12.1%	0.5%	0.0%
Median Fox Confidence	0.71	0.87	0.85	0.89
False Fires on 225 Clean Negatives	13 (5.8%)	2 (0.9%)	4 (1.8%)	0 (0.0%)

Operating Threshold Guide

• Recommended Operating Point (conf = 0.40): Optimum balance for automated alerts. Delivers 88.2% fox capture with zero rival canid false triggers.
• Search / Discovery (conf = 0.25): Captures 88.5% of frames with minimal low-confidence background noise.
• Strict High-Precision Alarm (conf = 0.50): Zero false positives across any lighting condition with ~78% frame capture.

---

Quantization & Edge Compression

TinyFox-2 includes INT8 calibrated weights generated with YOLO-Quantizer (per-channel Conv weight quantization with sensitivity analysis).

Format	File	Size	Reduction	Mean Box Delta	Class Agreement
PyTorch (FP32)	TinyFox-2.pt	5.39 MB	—	Baseline	100%
ONNX (FP32)	TinyFox-2.onnx	9.48 MB	—	Baseline	100%
ONNX (FP16)	TinyFox-2_fp16.onnx	4.75 MB	50.0%	< 0.05 px	100%
ONNX (INT8)	TinyFox-2_int8.onnx	2.60 MB	72.5%	0.68 px	100%

On validation images, the INT8 model matches the FP32 model with an average bounding box shift of less than 0.7 pixels on a 640×640 grid and 100% class match.

---

Repository Files

TinyFox-2/
├── README.md                      # Model Card & Documentation
├── LICENSE                        # GNU AGPL-3.0 + Attribution Notice
├── config.json                    # Architecture & class metadata
├── TinyFox-2.pt                   # PyTorch model checkpoint (Ultralytics)
├── TinyFox-2_last.pt              # Resume training checkpoint
├── TinyFox-2.onnx                 # FP32 ONNX model (opset 20, 640x640)
├── TinyFox-2_fp16.onnx            # FP16 ONNX model (for GPU/NPU)
├── TinyFox-2_int8.onnx            # Calibrated INT8 ONNX (2.60 MB)
├── TinyFox-2.torchscript          # TorchScript traced model
├── infer.py                       # Ultralytics inference script
├── infer_onnx.py                  # Standalone ONNX Runtime runner (no PyTorch required)
├── export_all_formats.py          # Multi-format export utility
├── yolo_quantizer.py              # YOLO mixed-precision quantizer (MIT)
├── upload_to_hf.ps1               # Hugging Face upload script
└── evaluation/                    # Performance curves & validation plots
    ├── BoxF1_curve.png
    ├── BoxPR_curve.png
    ├── confusion_matrix.png
    ├── results.png
    └── results.csv

---

Quickstart & Usage

1. PyTorch / Ultralytics API

from ultralytics import YOLO

# Load model
model = YOLO("TinyFox-2.pt")

# Predict on an infrared image
results = model.predict(source="night_frame.jpg", conf=0.40, imgsz=640)

# Display or save results
for r in results:
    r.show()  # Display
    r.save(filename="output.jpg")  # Save

2. Command-Line Inference

# Run on an image
python infer.py --source test_ir.jpg --conf 0.40 --output result.jpg

# Run on a video file
python infer.py --source trail_cam.mp4 --conf 0.40 --output annotated.mp4

# Run on an RTSP stream (live monitor)
python infer.py --source "rtsp://192.168.1.100:554/stream" --conf 0.40 --show

3. Lightweight ONNX Runtime (Zero PyTorch Dependency)

For low-power edge nodes (Raspberry Pi, mini PCs) without PyTorch:

pip install onnxruntime opencv-python numpy
python infer_onnx.py --model TinyFox-2_int8.onnx --image test_ir.jpg --conf 0.40

---

Exporting & Quantizing

To re-quantize or export to additional formats:

# Quantize to INT8 with YOLO-Quantizer
python yolo_quantizer.py TinyFox-2.onnx --mode int8 --per-channel --out TinyFox-2_int8.onnx

# Export to TorchScript / OpenVINO
python export_all_formats.py --model TinyFox-2.pt --imgsz 640

---

Citation & Acknowledgments

If you use TinyFox-2 in your research or deployment, please cite:

@misc{tinyfox2026,
  title={TinyFox-2: High-Performance Night/IR Red Fox and Wildlife Object Detector},
  author={Thenukegun10x and Contributors},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/Thenukegun10x/TinyFox-2}}
}

Upstream Acknowledgments

• Ultralytics YOLO: https://github.com/ultralytics/ultralytics (AGPL-3.0).
• NTLNP Dataset: Northeast China Tiger and Leopard National Park research group, Beijing Normal University.
• UNSW Dryland Ecology: Centre for Ecosystem Science, University of New South Wales.
• LILA BC (Camera Traps): Eastern North America 2024 (ENA24) and Snapshot USA contributors.
• Microsoft AI for Earth: MegaDetector v6 camera-trap framework.
• YOLO-Quantizer: https://github.com/Thenukegun10x/YOLO-Quantizer (MIT).