thenukegun10x/TinyFox-1.0

TinyFox 1.0

TinyFox 1.0 is a tiny one-class red fox detector trained for night/IR camera-trap footage.

The model is based on a lightweight NanoDet-style detector with multi-scale heads at strides 8/16/32. It predicts one class: red_fox.

Files

hf_tinyfox_1_0/
├── TinyFox_1-0.pt              PyTorch state_dict checkpoint
├── tinyfox_config.json         architecture & preprocessing metadata
├── export_all_formats.ps1      export script
├── upload_to_hf.ps1            Hugging Face upload helper
├── README.md                   this file
├── onnx/
│   ├── TinyFox_1-0_fp32.onnx   FP32 ONNX (safest generic export)
│   ├── TinyFox_1-0_fp16.onnx   FP16 ONNX (GPU/NPU runtimes)
│   └── TinyFox_1-0_int8.onnx   INT8 dynamic-quantized ONNX
├── safetensors/
│   ├── TinyFox_1-0_bf16.safetensors  BF16 weights archive
│   └── TinyFox_1-0_fp16.safetensors  FP16 weights archive
├── torchscript/
│   └── TinyFox_1-0_torchscript.pt    TorchScript trace
├── tflite/
│   └── TinyFox_1-0.tflite            TFLite (float32 + float16)
├── gguf/
│   └── TinyFox_1-0_fp16.gguf         GGUF weights archive
└── acuity/                            ACUITY NPU bundle
    ├── nanodet.onnx
    ├── nanodet_acuity.json
    ├── README_ACUITY.txt
    └── calibration_images.txt

Metrics

Validation after Frigate hard-negative refinement:

Metric	Value
mAP@0.50	0.9925
Recall@0.50	1.0000
Precision@eval threshold	0.0558
Mean matched IoU	0.8819
Validation loss	1.3707

The precision value above is measured with a low evaluation threshold for mAP-style scoring and should not be interpreted as the recommended deployment threshold.

Intended Use

• Red fox detection in night/IR camera footage.
• Edge/NPU deployment experiments.
• Wildlife monitoring pipelines where zoning and post-processing can be used to reduce static-scene false positives.

Recommended inference confidence range:

• Start with 0.25.
• Lower toward 0.15 if misses are unacceptable.
• Raise toward 0.35 if false positives are unacceptable.

For static cameras, use zones to suppress detections in impossible/irrelevant areas such as timestamp overlays, sky, fixed leaf piles, or walls.

Limitations

• One class only: red fox.
• Optimized for night/IR footage, not broad daytime wildlife detection.
• Validation contains pseudo-labeled and project-specific data, so metrics may overestimate general-world accuracy.
• Post-processing is required for ONNX/ACUITY raw outputs.
• GGUF is included only as a weights archive format, not as a standard detector runtime.

PyTorch Usage

From the original project repository:

python -m foxdetect.infer_nanodet --checkpoint hf_tinyfox_1_0\TinyFox_1-0.pt --image test.jpg --output result.jpg --conf 0.25

For video:

python -m foxdetect.infer_nanodet --checkpoint hf_tinyfox_1_0\TinyFox_1-0.pt --video input.mp4 --output result.mp4 --conf 0.25

Architecture parameters:

img_size = 512
width = 1.0
reg_max = 16
strides = 8, 16, 32
classes = red_fox

Preprocessing

Input image preprocessing:

RGB
resize to 512 x 512
scale to 0..1
normalize mean = [0.485, 0.456, 0.406]
normalize std  = [0.229, 0.224, 0.225]
layout = NCHW

ONNX Outputs

The ONNX export returns raw detector head tensors:

cls_s8, reg_s8
cls_s16, reg_s16
cls_s32, reg_s32

Post-processing steps:

1. Apply sigmoid to class logits.
2. Decode DFL regression with reg_max=16.
3. Use strides 8,16,32.
4. Scale decoded boxes back to the source image size.
5. Apply thresholding and NMS.

Export Commands

Run from the repository root:

.\hf_tinyfox_1_0\export_all_formats.ps1

Individual commands:

python -m foxdetect.export_nanodet --checkpoint hf_tinyfox_1_0\TinyFox_1-0.pt --format onnx --dtype fp32 --output hf_tinyfox_1_0\onnx\TinyFox_1-0_fp32.onnx
python -m foxdetect.export_nanodet --checkpoint hf_tinyfox_1_0\TinyFox_1-0.pt --format onnx --dtype fp16 --output hf_tinyfox_1_0\onnx\TinyFox_1-0_fp16.onnx
python -m foxdetect.export_nanodet --checkpoint hf_tinyfox_1_0\TinyFox_1-0.pt --format onnx-int8 --output hf_tinyfox_1_0\onnx\TinyFox_1-0_int8.onnx
python -m foxdetect.export_nanodet --checkpoint hf_tinyfox_1_0\TinyFox_1-0.pt --format safetensors --dtype bf16 --output hf_tinyfox_1_0\safetensors\TinyFox_1-0_bf16.safetensors
python -m foxdetect.export_nanodet --checkpoint hf_tinyfox_1_0\TinyFox_1-0.pt --format safetensors --dtype fp16 --output hf_tinyfox_1_0\safetensors\TinyFox_1-0_fp16.safetensors
python -m foxdetect.export_nanodet --checkpoint hf_tinyfox_1_0\TinyFox_1-0.pt --format torchscript --output hf_tinyfox_1_0\torchscript\TinyFox_1-0_torchscript.pt
python -m foxdetect.export_nanodet --checkpoint hf_tinyfox_1_0\TinyFox_1-0.pt --format acuity --output hf_tinyfox_1_0\acuity
python -m foxdetect.export_nanodet --checkpoint hf_tinyfox_1_0\TinyFox_1-0.pt --format tflite --output hf_tinyfox_1_0\tflite\TinyFox_1-0.tflite
python -m foxdetect.export_nanodet --checkpoint hf_tinyfox_1_0\TinyFox_1-0.pt --format gguf --dtype fp16 --output hf_tinyfox_1_0\gguf\TinyFox_1-0_fp16.gguf

ACUITY Notes

Use the acuity/ export folder for Vivante/TIM-VX style NPU toolchains. The ACUITY bundle contains:

• nanodet.onnx
• nanodet_acuity.json
• README_ACUITY.txt
• calibration_images.txt

For INT8/INT4 deployment, prefer the vendor calibration flow from the FP32 ONNX model instead of generic ONNX INT4.

Upload To Hugging Face

huggingface-cli login
huggingface-cli repo create TinyFox-1.0 --type model
.\hf_tinyfox_1_0\upload_to_hf.ps1 -RepoId YOUR_USERNAME/TinyFox-1.0

Training Data Summary

The detector was trained from a combined night-focused dataset including:

• Real red fox night/camera-trap images.
• LocateAnything-assisted UNSW night red fox pseudo-labels.
• Hard negatives from non-fox VOC night animals.
• Frigate false-positive negatives from deployment-like IR footage.

Hard negatives were important for reducing false positives on branches, leaves, IR glare, and static bright ground regions.