No Clear Winner in NIST Age Estimation Assessment

“There is a wide range in performance among these algorithms, with room for improvement across the board.” – NIST computer scientist Kayee Hanaoka

The National Institute of Standards and Technology (NIST) has released its first report in a decade on the performance of software algorithms that estimate a person’s age from a photo. NIST evaluated six algorithms and found none clearly superior to the others. The agency plans to update the evaluation results every four to six weeks, expecting improvements in the software’s capabilities due to advancements in AI and biometric tech.

Titled “Face Analysis Technology Evaluation: Age Estimation and Verification (NIST IR 8525),” the study assessed algorithms submitted voluntarily in response to a 2023 call. The results showed varied performance, with all algorithms leaving room for improvement.

The FATE study marks the beginning of a new, long-term effort by NIST to regularly test age estimation and verification (AEV) technology.

NIST’s last evaluation of AEV software was in 2014, using a single database of photos from visa applications. Since then, interest in the technology has grown, prompting NIST to expand its evaluation program. The new study used approximately 11.5 million photos from diverse databases, including FBI mug shots, webcam images from border crossings, and immigration application photos, all anonymized to protect privacy.

NIST’s researchers found no single algorithm excelled in all areas, with accuracy influenced by factors such as image quality, gender, region of birth, and age. Algorithms showed varying sensitivities to different demographic groups. Despite improvements since 2014, error rates remained higher for female faces than males, a consistent trend from the previous evaluation.

NIST’s ongoing testing program aims to release updates every four to six weeks. The team plans to refine their test methods and expand their photo databases to cover more applications, like online safety. Future evaluations might include additional criteria, such as using prior photos of the same person to improve accuracy.

Source: NIST

–

May 30, 2024 – by Cass Kennedy