How Alibaba Optimizes Campus‑Scale Face Recognition for 200K Daily Scans

Alibaba’s campus uses face recognition for access control, payment, and parcel pickup, processing over 200,000 scans daily.

Yang Hanfei (alias Shaohao), head of the Park Brain tech team, notes the first facial gate was deployed in January, and now the system supports large events like the Cloud Expo with 120,000 attendees.

How the system works

Face registration requires a photo; during recognition the terminal captures an image, extracts features, and compares them with the server‑side algorithm to verify identity.

Terminal Optimization

Hardware improvements include wide‑dynamic‑range cameras and IPS optimization to handle back‑lighting and capture clear foregrounds.

Software manages image clarity by monitoring each frame, labeling data, and classifying images by sharpness and angle, which controls recognition speed and distance.

Algorithm Optimization

Reducing false‑accept rates and managing similarity thresholds are key; as the face database grows, algorithm performance can degrade, so predictive models and grouping optimization are applied.

Grouping optimization splits the population by time and space, creating smaller cohorts to lower mis‑recognition; it also incorporates business rules and historical behavior to handle parcel‑pickup delegation.

These models have brought the mis‑recognition rate down to about 0.1% (1‰).

Database Photo Optimization

Poor quality enrollment photos cause failures; the team uses algorithms for size detection, correction, and identity verification, and employs self‑learning to cluster and assess photo quality.

During the 2022 Double 11 shopping festival, the same face‑recognition system protected the command center and the live‑stream venue.

Future plans aim to enrich facial attribute recognition, improve hardware and algorithms, and expand to more commercial scenarios, heralding a true “face‑scan” era.