ONNX + YOLO Face Analytics

A playful weekend computer-vision experiment using ONNX and YOLO-style detection to put face analytics over video. The fun bit is seeing a practical pipeline come together: detection boxes, per-face annotations, and a light overlay that makes the model behavior visible frame by frame.

The prototype is intentionally rough. The gender detection is still a little screwy, and that part needs further work before it should be treated as reliable. For now, it is best read as a hands-on systems experiment rather than a finished analytics product.

Technical note

The useful learning is in the plumbing: model export/runtime behavior, frame processing, detector confidence, annotation layout, and how quickly small perception errors become obvious once they are rendered on top of moving footage.

Why this is interesting

Computer vision demos become much more honest when the predictions are drawn directly onto video. You can see jitter, missed detections, classification uncertainty, and whether the model is stable enough for real use.

Next work

• Improve gender-classification reliability
• Tune confidence thresholds
• Check bias and failure cases
• Separate detection from downstream analytics