In the FRVT report released by NIST on March 04, 2021, the IREX’s face recognition algorithm was ranked in top 10 in accuracy for these datasets among 268 algorithms submitted by companies around the world.
Border and Kiosk Photos are the most relevant datasets to identify missing people and suspects in non-cooperative mode from arbitrary city cameras. These datasets are less constrained in terms of viewing angle, lighting, and resolution when compared to the other FRVT datasets.
The following table gives at left, rank 1 miss rates relevant to investigations; at right, with threshold set to target FPIR = 0.01 for higher volume, low prior, uses. Yellow indicates the most accurate algorithm.
Smaller is better.
The IREX face recognition algorithm is highly optimized for non-cooperative recognition in real-world CCTV systems. While the FRVT estimates the recognition accuracy on still images, IREX uses an additional layer of neural networks to create aggregated templates from multiple video frames. The advanced technique further increases the real-world accuracy.
The IREX algorithm also detects various facial features including; gender, age, glasses, beard, mustaches, face mask, headdress, and skin color. IREX can generate alerts for safety regulation violations such as the lack of face mask, protective glasses or helmet.
The IREX algorithm does identify people wearing masks. According to the NISTIR 8331 - Ongoing FRVT Part 6B: Face recognition accuracy with face masks using post-COVID-19 algorithms, IREX ranked #63 of 198 algorithms evaluated to date.
Watch the IREX face recognition demo to find missing children.