Vision Pipelines & Transfer Learning

Modern CV systems reuse pretrained visual representations.

Pipeline

1. Define the decision (classify, detect, segment)
2. Curate and augment data without leaking test signal
3. Start from a pretrained backbone
4. Fine-tune heads, then deeper layers if needed
5. Evaluate with task-appropriate metrics

Pitfalls

• Train/val contamination via augmentation or duplicates
• Ignoring domain shift (medical, industrial, satellite)
• Over-trusting accuracy on imbalanced classes