Start with the question, not the model
A face workflow should first state what it needs to learn. Detection answers where a face is. Alignment, cropping, and landmarks make the face region more consistent. None of these steps identify a person.
Keep comparison and recognition separate
Descriptors are machine-readable representations of detected faces. A 1:1 comparison measures two images against a threshold, while gallery recognition searches one target against labeled examples. A distance, boolean match, or gallery candidate is evidence for review, not proof of identity. Lighting, pose, occlusion, image quality, and the chosen threshold can change the result.
Treat attributes as estimates
Expression tools classify visible cues, and age-gender tools estimate attributes from an image. These outputs are not ground truth about emotion, age, gender, or a person’s character. Record the uncertainty and avoid using them as the sole basis for employment, access, education, housing, financial, law-enforcement, or other high-impact decisions.
Put privacy and error review in the workflow
Use images only with a lawful purpose and appropriate notice or consent. Minimize collection, protect images and descriptors, set a retention period, and delete them when the purpose ends. Test representative conditions for false positives and false negatives, route uncertain cases to a person, and document the threshold and review outcome before delivering a result.