Applications & Ethics

Computer Vision in the Real World

🔄 Lessons 2-6 covered the technical foundations — how images are represented, how CNNs detect patterns, how detection and segmentation work, and how transfer learning makes it practical. Now let’s see where these techniques are deployed at scale, and the ethical challenges they create.

Autonomous Vehicles

Self-driving systems combine multiple CV models running simultaneously:

• Object detection: Identify cars, pedestrians, cyclists, traffic signs
• Semantic segmentation: Map drivable road vs sidewalk vs barrier
• Depth estimation: Judge how far away objects are
• Lane detection: Find lane markings and road boundaries
• Motion prediction: Anticipate where detected objects will move next

Waymo’s sixth-generation system uses cameras and lidar with AI to achieve a 360-degree field of view, identifying objects up to 500 meters away — day, night, and in poor weather. The system processes dozens of camera feeds simultaneously at 30+ FPS.

The challenge: Edge cases. The system handles 99.9% of driving scenarios. The remaining 0.1% — construction zones with confusing signage, emergency vehicles approaching from unexpected angles, objects that don’t fit any training category — is where accidents happen.

Medical Imaging

Computer vision has transformed diagnostic imaging:

• Radiology: CNN-based systems detect tumors, fractures, and abnormalities in X-rays, CT scans, and MRIs with accuracy matching or exceeding radiologists for specific tasks
• Pathology: Whole-slide imaging + deep learning classifies tissue samples for cancer staging
• Ophthalmology: Retinal scans processed by CV models detect diabetic retinopathy, glaucoma, and macular degeneration
• Dermatology: Smartphone-based systems classify skin lesions with dermatologist-level accuracy

The deployment model: AI assists, not replaces. CV systems triage scans (flagging suspicious findings for priority review), provide second opinions, and catch subtle findings that might be missed during high-volume reading sessions. The radiologist makes the final diagnosis.

✅ Quick Check: A radiology AI system achieves 96% accuracy on detecting pneumonia — matching expert radiologists. Should it make diagnoses autonomously? No. 96% accuracy means 4% errors. On 1,000 scans per day, that’s 40 wrong diagnoses — some potentially fatal false negatives. The system also can’t explain its reasoning, handle edge cases, or consider the patient’s full clinical context. The right deployment: AI flags suspicious scans for priority review, reducing workload while keeping radiologists in the decision loop. Human-AI collaboration typically outperforms either alone.

Manufacturing Quality Control

CV-based inspection systems are deployed across manufacturing:

• Surface defect detection: Cameras identify scratches, dents, discoloration, and micro-cracks invisible to human inspectors
• Dimensional measurement: CV systems measure component dimensions to sub-millimeter precision
• Assembly verification: Confirm all components are present and correctly positioned
• Predictive maintenance: Multi-angle vision systems identify equipment wear before failure, reducing downtime by up to 40%

Manufacturing leads CV adoption (35-37% of the market) because the environment is controlled: consistent lighting, known product shapes, fixed camera positions, and clear success criteria (defect or no defect).

Retail and Security

Retail: Cashierless checkout (Amazon Go), inventory management (shelf monitoring), customer behavior analysis, product recognition at point-of-sale.

Security: Surveillance analytics, anomaly detection (unattended bags, crowd density), access control, and perimeter monitoring.

Ethical Challenges

Bias in Computer Vision

CV systems inherit and amplify biases from training data. Documented examples:

• Facial recognition accuracy drops up to 40% for underrepresented demographics
• Medical imaging models trained primarily on one population perform poorly on others
• Object detection models trained on Western image datasets misidentify objects common in other cultures

The root cause: Training data skews. ImageNet, COCO, and most benchmark datasets overrepresent certain demographics, geographies, and contexts. Models perform best on data similar to their training distribution and worse on underrepresented groups.

Deepfakes and Synthetic Media

Deepfake technology creates convincing fake images, videos, and audio. A single deepfake CEO video cost one company $25.6 million in wire fraud. Detection methods achieve 90-95% accuracy but face an arms race — as detection improves, generation improves too.

Current approaches to the deepfake problem:

• Detection models — spot artifacts, but struggle with novel generation methods
• Provenance tracking (C2PA) — cryptographic proof of content origin
• Digital watermarking — embed invisible markers in authentic content
• Platform policies — label synthetic content, restrict distribution

Surveillance and Privacy

The technical capability for mass visual surveillance now exists. Facial recognition can identify individuals in crowds. Gait analysis can identify people even when faces are obscured. License plate readers track vehicle movements. The question isn’t whether these technologies work — it’s whether and how they should be used.

The EU AI Act restricts real-time biometric identification in public spaces. Several US cities ban government facial recognition. But the technology continues to advance faster than policy.

✅ Quick Check: A company wants to deploy facial recognition for employee access control. What ethical considerations should they address? Consent (employees must opt in, with a non-biometric alternative available), data security (biometric templates are permanently compromising if breached — you can’t change your face), bias testing (verify accuracy across all employee demographics), data retention (how long is biometric data stored, and who has access), and legal compliance (GDPR, BIPA, and other biometric privacy laws).

Key Takeaways

• Autonomous vehicles combine detection, segmentation, depth, and motion prediction — handling 99.9% of scenarios, struggling with edge cases
• Medical imaging CV assists radiologists rather than replacing them — human-AI collaboration outperforms either alone
• Manufacturing leads CV adoption (35-37%) due to controlled environments and clear success criteria
• Bias in training data causes accuracy drops of up to 40% for underrepresented groups
• Deepfake detection faces an arms race — provenance tracking (C2PA) is more sustainable than detection alone
• Surveillance CV creates mass biometric collection — scanning everyone to find the few

Up Next

You understand the technology, applications, and risks. Lesson 8 brings it all together — designing your first project, choosing a career path, and mapping the skills that command $128K-$208K salaries.