Vision AI Attendance

Face-based attendance needs to solve more than simply detecting a face. The initial prototype relied on a single reference image, repeated an expensive CNN detection step during alignment, and displayed a similarity value that could easily be mistaken for model accuracy. These limitations increased processing cost and made identity matching less robust when the user's pose or capture condition changed.

I audited the complete computer-vision pipeline and measured individual processing stages before changing the architecture. A redundant CNN face-detection pass during alignment was removed through an optimized FAST_ALIGN path, and the original single-image registration flow was replaced with guided multi-sample enrollment across center, left, and right poses. The recognition pipeline now supports multiple reference embeddings per identity while remaining backward-compatible with legacy users. I also added enrollment quality gates, benchmark instrumentation, automated tests, and threshold-calibration tooling so matching and performance decisions can be evaluated from measured evidence rather than arbitrary percentages.

Vision AI V2 provides an end-to-end attendance workflow from guided face registration to identity matching, randomized liveness verification, duplicate-attendance prevention, and administrative analytics. Live development testing showed that removing redundant alignment detection substantially reduced processing cost in that stage, while multi-reference enrollment allows each identity to be represented across multiple poses without meaningful matching overhead at the current project scale. The remaining limitations of the system, including CPU cost, heuristic passive spoof detection, and limited threshold-calibration data, are documented explicitly instead of being hidden behind unsupported accuracy claims.