yolov8-piglife-sam3

YOLOv8 Models for Pig Monitoring (Trained on SAM3 Pseudo-Labels)

This repository provides the official YOLOv8 weights (Nano, Small, Medium) presented in our paper:

SAM3-Assisted Training of Lightweight YOLO Models for Precision Pig Farming
Marcos Vinicius Mendes Faria, Thiago Borges Pereira, Isabella C. F. S. Condotta, Thiago Meireles Paixão, and Francisco de Assis Boldt. (2026)
arXiv: arXiv:2605.25860 | GitHub: mvmfaria/sam3-pseudo-label-pig-yolo

These models were trained entirely on pseudo-labels generated by the Segment Anything Model 3 (SAM3) applied to the PigLife dataset. They offer an off-the-shelf solution for precision livestock farming without needing to run heavy foundation models in production.

[!TIP]

Training Dataset

If you wish to train your own models from scratch, we have made the SAM3 pseudo-labels publicly available. Download the Dataset (Labels) Here

<p align="center"> <img src="cover.png" alt="SAM3 Training Pipeline" width="100%"/> </p>

Trained YOLOv8 Models

These models are lightweight and optimized for real-time inference on edge devices (like Raspberry Pi or Jetson Nano) as well as servers.

Model Weights Availability

• yolov8n-pig.pt (Nano) - Best for edge devices.
• yolov8s-pig.pt (Small) - Good balance of speed and accuracy.
• yolov8m-pig.pt (Medium) - Highest accuracy for server-side processing.

Quick Start (Inference)

You can use our trained models directly via the ultralytics package by passing the Hugging Face URL:

from ultralytics import YOLO

# Load the trained model directly from Hugging Face
model = YOLO("hf://marcosvmf/yolov8-piglife-sam3/yolov8s-pig.pt")

# Run inference on an image or video
results = model.predict("path/to/your/pig_image.jpg", save=True)
results[0].show()

---

Performance Baselines (COCO Metrics)

Object detection performance on the test set comparing models trained on Human Annotations vs. SAM 3 Pseudo-labels, as well as the SAM 3 Zero-shot baseline:

Key takeaway: Models trained on SAM 3 pseudo-labels retain up to 98.4% of SAM 3's zero-shot detection performance (79.4 vs 80.7 mAP, and identical 93.6% AP₅₀) while reducing latency from ~1,242 ms to 10–16 ms (over 70× – 120× speedup).

---

Limitations

• Domain Shift: The models are optimized for the camera angles and lighting conditions present in the PigLife UIUC dataset. Performance may vary in vastly different farm setups.
• Pseudo-Label Noise: Since the training labels were generated by a foundation model (SAM3) and not human experts, some edge cases might be incorrectly learned by this model.

---

Citation

If you use these models or pseudo-labels in your research, please cite our paper:

@misc{faria2026sam3assistedtraininglightweightyolo,
      title={SAM3-Assisted Training of Lightweight YOLO Models for Precision Pig Farming}, 
      author={Marcos Vinicius Mendes Faria and Thiago Borges Pereira and Isabella C. F. S. Condotta and Thiago Meireles Paixão and Francisco de Assis Boldt},
      year={2026},
      eprint={2605.25860},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.25860}, 
}