YOLO Pipeline

When to Use This Skill

• Detecting objects with bounding boxes in images or video
• Segmenting objects at pixel level
• Classifying entire images into categories
• Estimating human body pose (keypoints)
• Training custom YOLO models on domain-specific data
• Evaluating model performance with IoU/mAP metrics

Decision Framework

Which YOLO Task?

What output do you need?
├── "What objects are here and WHERE?" (bounding boxes)
│   └── ✅ Detection — model('img.jpg') with yolov8n.pt
│
├── "Exact pixel-level shape of each object"
│   └── ✅ Segmentation — model('img.jpg') with yolov8n-seg.pt
│
├── "What IS this image overall?" (single label)
│   └── ✅ Classification — model('img.jpg') with yolov8n-cls.pt
│
├── "What pose is this person in?" (17 keypoints)
│   └── ✅ Pose Estimation — model('img.jpg') with yolov8n-pose.pt
│
└── "Rotated/angled objects" (ships, aircraft)
    └── ✅ Oriented Bounding Boxes (OBB) — yolov8n-obb.pt

Semantic vs Instance Segmentation

| Type | Question | Output | Can Count Individuals? | |------|----------|--------|------------------------| | Semantic | "What’s in the scene?" | One color per class | ❌ No (5 sheep = one green blob) | | Instance (YOLO) | "Which objects where?" | Unique ID per object | ✅ Yes (5 sheep = 5 different colors) |

> YOLO does Instance Segmentation — each object gets its own mask and identity.

Pre-trained Dataset Reference

| Dataset | Source | Classes | Typical Use | |---------|--------|---------|-------------| | COCO | Microsoft | 80 | General object detection | | ImageNet | Stanford | 1000 | Image classification | | DOTAv1 | Wuhan Univ. | 15 | Aerial/satellite OBB |

When Classic CV Fails → Use Deep Learning

| Condition | Classic CV | YOLO/DL | |-----------|-----------|----------| | High contrast, uniform light | ✅ Works | Overkill | | Shadows, uneven lighting | ❌ Fails | ✅ Robust | | Touching/overlapping objects | ❌ Fails | ✅ Handles | | Diverse viewpoints | ❌ Unreliable | ✅ Generalizes | | Need to detect 80+ classes | ❌ Impractical | ✅ Built-in |

YOLO Detection — The 3-Step Pipeline

Step 1: Grid Division (Localization)

• Image is divided into grid cells (e.g., 7×7)
• Each cell is responsible for detecting objects whose center falls within it
• "You Only Look Once" — each cell handles its own region

Step 2: Anchor Box Prediction

• Each cell generates Anchor Boxes of different aspect ratios
• For each anchor: confidence score + class probabilities + bbox coordinates
• YOLO predicts: [center_x, center_y, width, height] (normalized 0-1)

Step 3: Non-Maximum Suppression (NMS)

• Multiple overlapping boxes around the same object → keep only the best one
• Low-confidence boxes are filtered first
• Among remaining overlapping boxes, NMS keeps the highest-confidence box

Training Workflow

Data Preparation Checklist

1. Collect images — diverse angles, lighting, backgrounds
2. Label data — Tools: makesense.ai (free, web-based), CVAT, Roboflow
3. YOLO label format (one .txt per image):

[class_index] [center_x] [center_y] [width] [height] Example: 1 0.45 0.60 0.30 0.40 All coordinates normalized to 0-1 range.

1. Classification: no labels needed — folder name IS the label

dataset/ ├── train/ │ ├── COVID/ ← folder name = class label │ └── Normal/ └── val/ ├── COVID/ └── Normal/

1. Data Augmentation — use Albumentations (auto-adjusts bbox/polygon coords)

Key Training Parameters

| Parameter | What It Controls | Guidance | |-----------|-----------------|----------| | Learning Rate | Step size for weight updates | Too high = overshoots, too low = slow convergence | | Epochs | Full passes through dataset | More is not always better (overfitting risk) | | Batch Size | Images per training step | Limited by GPU memory. 16-32 typical | | imgsz | Input image resolution | 640 default. Higher = better accuracy, slower |

Transfer Learning

Don't train from scratch. Use pre-trained weights:

python
from ultralytics import YOLO
model = YOLO('yolov8n.pt')  # Pre-trained on COCO
model.train(data='my_data.yaml', epochs=100, imgsz=640)

> Transfer Learning = standing on giants' shoulders. The model already knows edges, textures, shapes from millions of images. You only retrain the final layers for your specific task.

Saved Weights

| File | Contains | Use For | |------|----------|--------| | best.pt | Weights at best validation accuracy | ✅ Production / inference | | last.pt | Weights at final epoch | Resume interrupted training only |

> Always use best.pt for inference. last.pt may have started overfitting.

Evaluation Metrics

Confusion Matrix Components

| | Predicted Positive | Predicted Negative | |---|---|---| | Actually Positive | TP (correct detection) | FN (missed detection — WORST in medical) | | Actually Negative | FP (false alarm) | TN (not used in object detection) |

> TN is not used in object detection systems — "correctly identifying nothing" is not meaningful.

IoU (Intersection over Union)

IoU = Area of Overlap / Area of Union

• Range: 0 (no overlap) to 1 (perfect match)
• IoU > 0.50 = detection counts as True Positive (standard threshold)
• Higher IoU thresholds = stricter evaluation

mAP (Mean Average Precision)

| Metric | Threshold | Strictness | |--------|-----------|------------| | mAP@0.50 | 50% overlap required | Standard | | mAP@0.50:0.95 | Average across 50%-95% | Very strict, comprehensive |

Precision, Recall, F1

Precision = TP / (TP + FP)    — "Of all detections, how many are correct?"
Recall    = TP / (TP + FN)    — "Of all real objects, how many did we find?"
F1        = 2 × (P × R) / (P + R)  — Harmonic mean

Overfitting Check

If training loss keeps dropping but validation loss plateaus or rises → overfitting. Solutions:

• More data augmentation
• Fewer epochs
• Regularization (dropout, weight decay)

XAI (Explainable AI) — EigenCAM

DL models are black boxes — they give answers but don't explain why.

| Heatmap Color | Meaning | |---------------|--------| | 🔴 Hot (Red/Yellow) | Model's decision focus — high attention | | 🔵 Cold (Blue/Purple) | Irrelevant to model's decision |

Why it matters in medicine:

• Verify the model looks at the lesion, not the hospital label in the corner
• Prove to clinicians that the AI decision is based on actual pathology
• Required for regulatory approval in many jurisdictions

Common Parameters

python
from ultralytics import YOLO

model = YOLO('yolov8n.pt')
results = model(
    source='image.jpg',    # or '0' for webcam, or video path
    show=True,             # display results
    save=True,             # save to disk
    conf=0.25,             # minimum confidence threshold
)

Industrial Lighting Systems

| Light Type | Setup | Best For | |-----------|-------|----------| | Array/Screen | Front-facing flat | Surface defects | | Ring Light | Around camera lens | Small parts, uniform light | | Back Light | Behind object | Silhouette, hole detection | | Bar Light | Angled beam | Directional surface inspection | | Dome Light | Surrounds object | Eliminating reflections on metal/shiny surfaces |