Researchers have found that face recognition algorithms can struggle to process identification when faced with massive datasets.A new report by a University of Washington team of researchers focuses on the so-called MegaFace Challenge, a system designed to evaluate how humans perform on sets of many hundreds of thousands of images versus computers.While previous test were performed on a dataset with only 13,000 images, the MegaFace dataset contains 1 million images representing more than 690,000 unique people.The testing leverages Yahoo's recently released database of Flickr photos. The Yahoo dataset includes 100 million creative commons photographs and hence can be released for research.”We need to test facial recognition on a planetary scale to enable practical applications – testing on a larger scale lets you discover the flaws and successes of recognition algorithms,” said Ira Kemelmacher-Shlizerman, a UW assistant professor of computer science and the project's principal investigator. “We can't just test it on a very small scale and say it works perfectly.”Facial recognition algorithms that fared well with 10,000 distracting images all experienced a drop in accuracy when confronted with 1 million images.Google's FaceNet showed the strongest performance on one test, dropping from near-perfect accuracy when confronted with a smaller number of images to 75 percent on the million person test. A team from Russia's N-TechLab came out on top on another test set, dropping to 73 percent.The goal was to test performance of face recognition algorithms with up to a million distractors, i.e., faces of unknown people. In each test, a “probe” image is compared against a gallery of up to a million faces drawn from the Megaface datasetThe researchers say the report had three key findings:1) The algorithms' performance degrades given a large gallery even though the probe set stays fixed.2) Testing at scale allows to uncover the differences across algorithms (which at smaller scale appear to perform similarly) 3) Age differences across probe and gallery are still more challenging for recognition.
Recent Articles
- Identity Week Europe 2026 raises the bar with 10 first-time speakers making their debut
- Sumsub joins World Economic Forum Unicorn Community to tackle AI fraud and advance inclusive digital identity
- Air carrier interface integrated into EES strategy
- Kentucky’s first mobile ID accepted for air travel
- FRONTEX expert to headline security document design at Identity Week Europe 2026
Recent Video
- New non-profit trade association of HAND will standardise global talent ID
- No cards, no phones: J.P. Morgan’s biometric revolution in payments
- Anthony Vidiano on building the future of access management
- Charisma Check wins 2025 Start-Up Pitching Competition, advancing identity verification and background checks for dating app users
- How to protect one billion Amazon customers: Lessons from continuous risk assessments
