YOLO11 Pill Detection Model
Model Description
This model is a custom-trained YOLO11 object detection model developed for detecting full pills in images.
It was trained using a labeled dataset prepared in Roboflow and exported in YOLO format.
The model identifies pills by predicting:
- Bounding boxes
- Confidence scores
- Object class labels
Currently, the model supports three detection classes:
This project is designed for pharmaceutical object detection research and computer vision experimentation.
Model Details
Model Type
- Architecture: YOLO11
- Task: Object Detection
- Framework: Ultralytics YOLO
Classes
1names:
2 0: pill
3 1: capsule
4 2: tablet
Intended Uses
Primary Use Cases
This model is intended for:
- Pill detection in images
- Pharmaceutical automation experiments
- Computer vision prototyping
- Medication localization in images/video
Out-of-Scope Use
This model is not intended for:
- Medical diagnosis
- Drug verification in clinical workflows
- Safety-critical pharmaceutical decisions
Predictions may be inaccurate under challenging imaging conditions.
Training Data
The model was trained on a
custom Roboflow dataset containing images of
Labeled_full_pill Computer Vision Model annotated with bounding boxes.
Dataset Characteristics
- Annotated in Roboflow
- Exported in YOLO8 format
- Single object class:
pill
Dataset Split
Example:
- Train: 70%
- Validation: 20%
- Test: 10%
Training Procedure
The model was trained using Ultralytics YOLO11 with pretrained weights.
Training Hyperparameters
1model: yolo11x.pt
2epochs: 100
3imgsz: 640
4batch: 16
5optimizer: auto
6device: 6
Training Command
1yolo train device=3 \
2 model=ul://ultralytics/yolo11/yolo11x \
3 data=ul://wijai-thongsom/datasets/labeled-full-pillv2iyolov8 \
4 roject=wijai-thongsom/jolly-husky \
5 name=yolo11x
6 epochs=100 \
7 imgsz=640 \
8 batch=-1
Evaluation Results
Model performance was evaluated on the validation set using standard object detection metrics.
Metrics
| Metric | Value |
|---|
| Precision | 0.953978 |
| Recall | 0.932336 |
| mAP50 | 0.965024 |
| mAP50-95 | 0.728589 |
Replace these values with the actual metrics from your training results.
Inference
CLI Example
1yolo detect predict \
2 model=yolo11x.pt \
3 source=image.jpg
Python Example
1from ultralytics import YOLO
2
3model = YOLO("yolo11x.pt")
4results = model("image.jpg")
5
6for result in results:
7 print(result.boxes)
Limitations
The model performance may degrade in cases such as:
- Poor lighting
- Motion blur
- Partial occlusion
- Overlapping pills
- Pill appearances not represented in the training dataset
Performance is dependent on image quality and dataset diversity.
Bias and Risks
Because this model was trained on a custom dataset, its predictions may be biased toward:
- Specific pill colors
- Particular lighting conditions
- Limited pill shapes and sizes
- Background styles present in training data
Use caution when applying the model to images outside the training distribution.
Environmental Impact
Training object detection models requires computational resources that consume energy.
Training setup example:
- Hardware: GPU
- Framework: Ultralytics YOLO11
- Epochs: 100
For reproducibility, document:
- GPU type
- Training duration
- Energy consumption estimate
Model Files
Typical files included in this repository:
1.
2├── README.md
3├── yolo11n.onnx
4├── yolo11m.onnx
5├── yolo11x.onnx
6├── data.yaml
7└── results.png
Citation
If you use this model, please cite:
1@misc{yolo11-pill-detection,
2 title={YOLO11 Pill Detection Model},
3 author={Wijai Thongsom},
4 year={2026},
5 publisher={Hugging Face}
6}
License
This model is released under the MIT License.
Acknowledgments
- Ultralytics for YOLO11
- Roboflow for dataset annotation/export
- Hugging Face Hub for model hosting
