Responsible AI
How VisionLab handles privacy, consent, bias and the limits of computer vision.
Privacy by design
Every module in VisionLab runs in your browser. Images, video frames and face templates are never uploaded, logged or stored. Models are downloaded once and cached locally.
Consent & purpose
Face features require explicit consent, blur faces by default, and avoid inferring sensitive attributes (age, gender, emotion, ethnicity). Real deployments need a lawful basis and clear signage.
Bias & fairness
Models inherit the biases of their training data (COCO, ImageNet, ADE20K, web image-text pairs). Measure error rates per subgroup and per environment before deploying.
Accuracy limits
Confidence is not certainty. Detection and verification errors have real costs; keep humans in the loop for consequential decisions and choose thresholds from the business cost of each error type.
Regulation
EU AI Act (remote biometric identification is high-risk or prohibited), GDPR Art. 9 on biometric data, Illinois BIPA, and sector rules such as FDA for medical imaging.
Transparency
Simulated figures (SPC history, FAR/FRR distributions, ROI defaults) are labelled as illustrative. Model names, sizes and sources are shown wherever a model runs.
Model cards (summary)
| Model | Trained on | Used for |
|---|---|---|
| SegFormer-B0 | ADE20K, 150 scene classes | Semantic segmentation |
| DETR ResNet-50 | COCO, 80 classes | Object detection |
| MobileViT-S | ImageNet-1k | Classification |
| CLIP ViT-B/32 | 400M web image–text pairs | Zero-shot labels, visual search |
| MediaPipe BlazeFace / Face Mesh / Hands | Face detection, landmarks, gestures |