---
license: other
license_name: bsl-1.1
license_link: https://mariadb.com/bsl11/
library_name: ultralytics
tags:
  - onnx
  - yolo
  - yolov8
  - object-detection
  - whiteboard
  - diagram
  - shapes
pipeline_tag: object-detection
---

# Whiteboard Detector

**Detects hand-drawn shapes on whiteboards.**

YOLOv8-nano fine-tuned to recognize 30 diagram shape classes.

## Quick Stats

| Spec | Value |
|------|-------|
| Architecture | YOLOv8-nano |
| Format | ONNX |
| Size | ~12 MB |
| Input | 640×640 RGB |
| Classes | 30 |
| Training | 100 epochs, 211 images |
| Hardware | M3 Max, 1.4 hours |

## Classes (30)

```
rectangle, rounded_rectangle, oval, circle, diamond, hexagon,
parallelogram, triangle, star, cloud, cylinder, stick_figure,
arrow_box, document_shape, database_icon, square, ellipse,
pentagon, cross, heart, lightning, banner, callout, bracket,
solid_arrow, dashed_arrow, bidirectional_arrow, dotted_line,
curved_arrow, curved_line
```

## Usage

### Python (ultralytics)

```python
from ultralytics import YOLO

model = YOLO("best.onnx")
results = model("whiteboard.jpg")

for box in results[0].boxes:
    cls = int(box.cls[0])
    conf = float(box.conf[0])
    x1, y1, x2, y2 = box.xyxy[0].tolist()
    print(f"{model.names[cls]}: {conf:.2f} at ({x1:.0f}, {y1:.0f})")
```

### Python (onnxruntime)

```python
import onnxruntime as ort
import numpy as np
from PIL import Image

# Load model
session = ort.InferenceSession("best.onnx")

# Preprocess
img = Image.open("whiteboard.jpg").resize((640, 640))
input_tensor = np.array(img).transpose(2, 0, 1).astype(np.float32) / 255.0
input_tensor = input_tensor[np.newaxis, ...]

# Inference
outputs = session.run(None, {"images": input_tensor})

# outputs[0] shape: [1, 34, 8400]
# 34 = 4 (xywh) + 30 (class scores)
# 8400 = detection candidates
```

### CLI (ultralytics)

```bash
yolo predict model=best.onnx source=whiteboard.jpg
```

## Output Format

YOLO outputs tensor `[1, 34, 8400]`:

```
For each of 8400 candidates:
  [0] x_center (0-640)
  [1] y_center (0-640)
  [2] width
  [3] height
  [4-33] confidence per class (30 classes)
```

Post-process with confidence threshold (0.25) and NMS (0.45 IoU).

## Training Performance

| Class | mAP50 | Notes |
|-------|-------|-------|
| cloud | 0.993 | Excellent |
| rounded_rectangle | 0.995 | Excellent |
| stick_figure | 0.895 | Good |
| oval | 0.849 | Good |
| rectangle | 0.716 | Good |
| text_label | 0.664 | Fair |
| solid_arrow | 0.368 | Needs more data |
| triangle | 0.316 | Needs more data |
| cylinder | 0.045 | Needs more data |

## Files

```
whiteboard-detector/
├── best.onnx        # Model (use this)
├── best.pt          # PyTorch weights
├── classes.txt      # Class names
├── README.md        # This file
└── SKILL.md         # Manifest
```

## Training Data

- 211 annotated whiteboard images
- Hand-drawn diagrams, varying styles
- Augmentation: rotation, blur, noise

## Limitations

- Best with clear contrast (dark ink on white)
- Small shapes (<20px) may be missed
- Overlapping shapes can confuse detection
- Some classes undertrained (cylinder, triangle)

## License

**Business Source License 1.1 (BSL-1.1)**

Copyright (c) 2024 Block Xaero Inc.

- ✅ Free for non-production use
- ⚠️ Production use requires license