Why this matters
Computer Vision Engineers ship models that affect people: identity verification, safety monitoring, retail checkout, medical imaging triage, and more. If error rates spike for certain groups (e.g., skin tones, age bands, assistive-device users), the product becomes unfair and risky. Regulators, customers, and your own QA teams expect evidence that you checked for bias and addressed it when found.
- Identity/face tasks: ensure similar detection/verification performance across demographic slices.
- Safety/PPE tasks: hardhats or reflective vests should be detected equally well on different body types, clothing colors, and lighting conditions.
- Autonomy tasks: pedestrian detection must work across mobility aids, strollers, and varying attire.
- Moderation/quality filters: acceptance rates should not systematically exclude certain groups.
Concept explained simply
Bias is a systematic error that disadvantages a group. Fairness checks are structured tests to quantify whether model performance is consistent across relevant slices of the population.
Mental model
Think of your data as a layered cake. Each slice (e.g., lighting=low, skin-tone=ST4–6, age=65+, camera=mobile) should have similar accuracy and error types. You probe each slice, compare metrics, and act if differences exceed small thresholds.
Key metrics you can compute quickly
- Selection Rate (SR): P(predicted positive). Used for Demographic Parity.
- True Positive Rate (TPR, recall): P(pred positive | actually positive).
- False Positive Rate (FPR): P(pred positive | actually negative).
- False Negative Rate (FNR): P(pred negative | actually positive).
- Equalized Odds: TPR and FPR are similar across groups.
- Demographic Parity: SR is similar across groups (regardless of ground truth).
- Calibration: predicted scores mean the same likelihood across groups.
- 80% (4/5) rule: SR_min / SR_max ≥ 0.8 is often used as a basic screening rule.
Typical tolerance examples: absolute differences ≤ 0.05 for TPR/FPR gaps; parity ratio ≥ 0.8 for SR. Choose thresholds appropriate to your domain and risk level.
A minimal workflow for fairness checks
- Define the decision: what counts as positive? what threshold?
- List slices that matter: e.g., skin tone bands, gender presentation, age bands, lighting, camera type, region.
- Ensure you have lawful basis to use any sensitive attributes. If not, consider proxies (lighting, device type) and consult your compliance team.
- Create a stratified evaluation set with enough examples per slice (aim ≥ 100 per slice to start; more is better).
- Compute baseline metrics overall: accuracy, precision, recall, FPR/FNR.
- Compute metrics per slice: SR, TPR, FPR, FNR, calibration curves if scores exist.
- Compare across slices: absolute differences, ratios (80% rule), and plots (e.g., ROC per group).
- Decide on mitigations if gaps exceed thresholds: targeted data collection, reweighting, augmentation, improved labeling, model changes, or calibrated thresholds.
- Re-test after mitigation. Document setup, metrics, and decisions.
- Schedule monitoring: run the same checks on new data regularly.
Ethics and compliance notes
- Use sensitive attributes only with proper consent/legal basis. Minimize who can access them and how long you store them.
- Per-group thresholds may improve fairness but can raise legal or policy questions. Document and get approvals before deployment.
- Avoid stereotyping or inferring sensitive attributes without a strong, approved reason.
Worked examples
Example 1 — Face detection rate by gender presentation
Data (ground-truth faces): Male-presenting: 1000 faces, model detects 960. Female-presenting: 1000 faces, model detects 910.
- TPR_male = 960/1000 = 0.96
- TPR_female = 910/1000 = 0.91
- Absolute gap = 0.05. If tolerance is ≤ 0.05, this is borderline. Action: investigate failure patterns; consider more training data representing hairstyles, makeup, occlusions.
Example 2 — Skin tone and false negatives (ST1–3 vs ST4–6)
Data (ground-truth faces): ST1–3: 800 faces, misses 40 → FNR = 40/800 = 0.05. ST4–6: 800 faces, misses 96 → FNR = 96/800 = 0.12.
- FNR gap = 0.07, which exceeds 0.05 tolerance.
- Likely causes: lighting imbalance, sensor noise, training data underrepresentation, exposure differences.
- Mitigations: collect low-light ST4–6 images; adjust exposure augmentation; verify labeling consistency; re-train and re-test.
Example 3 — PPE detection (hardhat) by lighting condition
Data: Daylight positives: 600, detected 558 → TPR_day = 0.93. Low-light positives: 600, detected 498 → TPR_low = 0.83.
- TPR gap = 0.10 (too high). Root cause: glare/ISO noise.
- Fixes: brightness/contrast augmentation, low-light synthetic noise, improved backbone or exposure-invariant preprocessing.
- Re-evaluate after mitigation aiming for gap ≤ 0.05.
Quick self-check checklist
- I defined which prediction is the “positive” decision and the threshold.
- I listed slices that matter for people and safety, not just what is easy to measure.
- I computed SR, TPR, FPR, FNR per slice and compared gaps.
- I applied at least one parity rule (e.g., 80%) or a clear gap tolerance.
- I documented findings and mitigation attempts.
- I planned monitoring to catch drift.
Hands-on exercises
Do these now. They mirror the exercises below, where you can check hints and solutions.
- Exercise 1: Compare TPR/FPR across two groups using provided confusion matrices and decide if equalized odds holds within 0.05.
- Exercise 2: Check the 80% rule for a quality filter’s acceptance rates and propose a mitigation.
Need a nudge? Mini task ideas
- Sketch a tiny table of metrics per slice (rows) and metric types (columns).
- Compute ratios twice: SR_min/SR_max and SR_max/SR_min to avoid mistakes.
- Sanity-check sums: TP+FN should equal positives; FP+TN equals negatives.
Common mistakes and how to self-check
- Too few examples per slice: Results bounce around. Self-check: add confidence intervals or increase sample size.
- Using only accuracy: Misses asymmetric errors. Self-check: always inspect TPR/FPR and FNR.
- Ignoring threshold effects: A single threshold may create gaps. Self-check: plot metrics vs threshold per group.
- Mixing label bias with model bias: Poor labels can mimic unfairness. Self-check: audit annotations across slices.
- One-and-done testing: Drift breaks fairness. Self-check: schedule periodic evaluation.
Practical projects
- Fairness audit notebook: Build a reusable notebook that ingests predictions and ground truth, then outputs per-slice SR, TPR, FPR, FNR, parity ratios, and a one-page summary.
- Data balancing pipeline: Create a small tool to detect underrepresented slices and propose sampling/augmentation settings to rebalance training data.
- Threshold explorer: Implement per-group metric curves (TPR/FPR vs threshold) to visualize fairness/accuracy trade-offs.
Who this is for
- Computer Vision Engineers and ML practitioners shipping models that interact with people or safety-critical workflows.
- QA/ML Ops folks who need to verify and monitor fairness over time.
Prerequisites
- Basic classification/detection metrics (precision, recall, confusion matrix).
- Comfort with Python or similar for metric computation.
- Awareness of your organization’s compliance policies for sensitive data.
Learning path
- Review confusion matrices and compute TPR/FPR/FNR reliably.
- Learn parity metrics (demographic parity, equalized odds, calibration).
- Practice slicing data and computing per-slice metrics.
- Apply mitigation techniques (data, model, thresholding) and re-test.
- Automate the checks and schedule monitoring.
Mini challenge
You have three slices with TPRs: 0.90, 0.88, 0.81. Your tolerance is 0.05. Which slice needs attention first, and what two quick mitigations would you try? Write your answer and compare it to your checklist.
Next steps
- Turn the checklist into a template for your team.
- Run the checks on your latest validation set and capture a one-page fairness report.
- Plan a re-run cadence (e.g., monthly) and who will review it.
About the quick test
The quick test is available to everyone. If you are logged in, your progress and score will be saved automatically.