marcosvmf/sam3-piglife-labels

SAM3 Pseudo-Labels for Pig Monitoring (Based on PigLife)

No Images Included (Copyright Restrictions)

Due to the copyright and licensing restrictions of the original PigLife dataset by UIUC, this repository DOES NOT contain any original images. This dataset provides only the pseudo-labels (annotations) generated by the Segment Anything Model 3 (SAM3) in YOLO format.

To use these annotations, you must legally acquire the original images directly from the official PigLife portal and pair them using our provided structure.

Pre-Trained YOLOv8 Models

If you don't want to train a model from scratch, we have published YOLOv8 models trained on this dataset. Download the Trained Models Here

Dataset Summary

This repository hosts the official pseudo-labels for 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

This project provides high-quality pseudo-labels for pig detection and segmentation, generated using SAM3 on the widely used PigLife dataset. Since manual annotation of dense livestock environments is prohibitively expensive, we leverage SAM3 to generate robust pseudo-masks and bounding boxes.

---

How to Reconstruct the Dataset

Since we cannot distribute the images, follow these steps to use our labels:

1. Download Original Images: Request and download the PigLife dataset from the official PigLife portal.
2. Download Our Labels: Clone this Hugging Face repository to get the YOLO-format .txt annotation files.
3. Merge Data: Place the downloaded images in an images/ folder and our labels in a labels/ folder. Ensure the filenames exactly match (e.g., frame_001.jpg needs to match frame_001.txt).

---

Pipeline: How SAM3 Pseudo-Labels Were Generated

1. Initial Prompts:
   • Foundation Model: We employed the open-vocabulary Segment Anything Model 3 (facebook/sam3) via Hugging Face transformers evaluated with torch.bfloat16 precision on GPU.
   • Text Prompting: The model was prompted zero-shot using the open-vocabulary text query "pig".
2. Refinement & Filtering:
   • Confidence Thresholding: Model predictions were post-processed through instance segmentation, filtering detections below a confidence threshold of $\tau = 0.40$ (CONFIDENCE_THRESHOLD = 0.4) to discard uncertain or noisy segmentations.
   • Intermediate COCO Representation: Filtered bounding box coordinates ([x_min, y_min, width, height]), areas, and categories were structured into standard COCO JSON format (instances_train.json, instances_val.json, instances_test.json).
3. Format Conversion:
   • COCO to YOLO: Using Ultralytics' dataset converter (convert_coco), annotations were converted into normalized YOLO TXT format: <class_id> <x_center> <y_center> <width> <height>.
   • Dataset Organization: Ground-truth files were categorized into respective train/, val/, and test/ splits accompanied by the official dataset.yaml.

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:

Supervision / Annotation	Model	Params (M)	Inf. Forward (ms)	Inf. Pipeline (ms)	mAP (50:95)	AP₅₀	AP₇₅	AP_M	AP_L
Human Annotated	YOLOv8n	3.2	5.70	9.29	88.1	98.9	95.7	67.5	89.0
Human Annotated	YOLOv8s	11.2	6.10	9.51	90.6	99.0	96.9	72.1	91.4
Human Annotated	YOLOv8m	25.9	14.70	15.43	91.7	99.0	97.8	75.2	92.5
SAM 3 Generated (Ours)	YOLOv8n	3.2	5.70	10.13	76.7	93.3	86.4	27.8	78.1
SAM 3 Generated (Ours)	YOLOv8s	11.2	6.10	10.37	78.7	93.5	87.7	32.1	80.1
SAM 3 Generated (Ours)	YOLOv8m	25.9	14.70	16.71	79.4	93.6	88.2	30.6	80.8
Zero-shot Baseline	SAM 3	~850	1197.18	1242.53	80.7	93.6	88.4	26.9	82.4

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

• Pseudo-Label Noise: The labels were generated by a foundation model (SAM3), not human experts. Some false positives or missed occluded pigs may exist.
• No Images: The dataset viewer will only show annotation metadata, not the images themselves.

---

Citation

If you use these pseudo-labels or models 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}, 
}