spotpredator-picodet

SpotPredator — PicoDet Farm Predator Detector (TFLite)

predator_picodet_s640_fp16.tflite is a fine-tuned PicoDet object-detection model that runs on-device on a Raspberry Pi Zero 2 W to detect common farm predators in real time. It is the vision model behind SpotPredator, a solar-powered, LoRa-linked field device that watches over free-range poultry and alerts a display station when a predator appears.

🔗 Full project (hardware, enclosure, code, deployment): github.com/JZVince/spotpredator

• Task: object detection
• Base model: PicoDet (PaddlePaddle / PaddleDetection), fine-tuned
• Format: TensorFlow Lite, FP16 quantized (fp16)
• Input: 640 × 640 RGB
• Runtime: tflite_runtime / ai-edge-litert on Raspberry Pi (CPU, XNNPACK)
• Detection classes: coyote, fox, raptor (predators only)
• Non-predators (background, people, poultry) are negative training instances — they carry no label and produce no detection.

Why PicoDet (was YOLO11n): the earlier SpotPredator model was fine-tuned from Ultralytics YOLO11 (AGPL-3.0). This model moves to PicoDet (Apache-2.0) for a permissive license — free to use, modify, and deploy commercially with attribution — with comparable on-device speed and improved precision (fewer false positives) after hard-negative retraining.

Intended use

Detecting farm predators (coyote, fox, raptor) from a fixed/rotating outdoor camera so a low-power edge device can trigger local alerts. Designed for low-resolution, small-object, edge-CPU conditions — predators often occupy only 20–40 px in a 1920×1080 frame, so the capture pipeline crops the sky band and tiles it into 640×640 patches before inference.

Out of scope / limitations

• Trained for a specific set of North-American farm predators; not a general wildlife detector.
• Small, distant, or heavily occluded animals may be missed.
• Performance varies with lighting, weather, and camera exposure.
• Not intended for safety-critical or human-detection use.

Training data

Fine-tuned on a mix of:

• Author-collected field images captured by the SpotPredator camera (Raspberry Pi Camera Module 3 / Arducam), representative of the real deployment (low-res, small objects, sky-cropped).
• LILA BC — Labeled Information Library of Alexandria: Biology and Conservation (camera-trap / wildlife imagery).
• GBIF — Global Biodiversity Information Facility occurrence media.

Please review and comply with the individual licenses/terms of the LILA and GBIF media used. Author-collected images are owned by the author.

Model output format

PicoDet is exported without in-graph NMS, so the TFLite model has two outputs:

• boxes: (1, N, 4) — decoded x1, y1, x2, y2 in input-pixel space (0–640), N ≈ 8500
• scores: (1, num_classes, N) — per-class confidence

You threshold + argmax the scores, filter boxes, and apply your own NMS.

Usage

# On a Raspberry Pi (or any TFLite host)
try:
    from tflite_runtime.interpreter import Interpreter
except ImportError:
    from ai_edge_litert.interpreter import Interpreter
import numpy as np
from PIL import Image

interpreter = Interpreter(model_path="predator_picodet_s640_fp16.tflite")
interpreter.allocate_tensors()
inp = interpreter.get_input_details()
out = interpreter.get_output_details()

# --- Preprocess: 640x640 RGB, /255 then ImageNet mean/std ---
MEAN = np.array([0.485, 0.456, 0.406], dtype=np.float32)
STD  = np.array([0.229, 0.224, 0.225], dtype=np.float32)
img = np.asarray(Image.open("frame.jpg").convert("RGB").resize((640, 640)), dtype=np.float32)
img = ((img / 255.0) - MEAN) / STD
interpreter.set_tensor(inp[0]['index'], img[None].astype(np.float32))
interpreter.invoke()

# --- Two outputs: identify boxes (ends in 4) vs scores by shape ---
box_i = 0 if out[0]['shape'][-1] == 4 else 1
boxes  = interpreter.get_tensor(out[box_i]['index'])[0]        # (N, 4)  x1,y1,x2,y2 in 0..640
scores = interpreter.get_tensor(out[1 - box_i]['index'])[0].T  # (N, num_classes)

labels = ["coyote", "fox", "raptor"]
CONF = 0.7                                                     # deployment threshold
cls_ids, confs = scores.argmax(1), scores.max(1)
for i in np.nonzero(confs >= CONF)[0]:
    print(labels[cls_ids[i]], round(float(confs[i]), 2), boxes[i].tolist())
# (apply your own NMS to de-duplicate overlapping boxes)

Confidence threshold used in the SpotPredator deployment: 0.7.

License

Apache-2.0. This model is fine-tuned from PicoDet (PaddlePaddle / PaddleDetection), which is licensed under Apache-2.0; derivative models inherit Apache-2.0. It is free to use, modify, and deploy — including commercially and in closed-source or networked services — with attribution. There is no copyleft/AGPL obligation.

Attribution

• Base model: PicoDet — PaddlePaddle/PaddleDetection (Apache-2.0)
• Datasets: LILA BC, GBIF, and author-collected imagery
• Project: SpotPredator — farm predator detection on the edge

Citation

@software{spotpredator_picodet,
  title  = {SpotPredator: PicoDet Farm Predator Detector (TFLite)},
  author = {JZVince},
  year   = {2026},
  url    = {https://huggingface.co/JZVince/spotpredator-picodet}
}