The most significant improvement is in face recognition accuracy, with as much as a 40% reduction in error rates in certain operating conditions.
In addition to Rank One’s standard face recognition algorithm, the ROC SDK also ships with an embedded algorithm designed for resource-constrained applications such as mobile devices or high-throughput video. While the accuracy for this algorithm has not changed relative to v1.19, it did receive a 10% speed improvement.
Other improvements delivered in v1.20 include:
- An overhaul of the face detection algorithm resulting in both accuracy and speed improvements
- The “Ethnicity” classifier has been changed to a “Geographic Origin” classifier, and can estimate if a person is of African, European, East Asian, or South Asian descent
- Profile faces (i.e., faces whose Yaw pose angle is greater than 60 degrees) can now be manually enrolled
- A new classifier for detecting facial drawings, which can flag face images that are not photorealistic
- Efficiency improvements to the (optional) CUDA GPU enrollment pipeline
- New licensing mechanisms, including the ability to sign a license file to a machine’s network “hostname”
- New API wrappers for Python 3 and the Lua programming language
The performance of the v1.20 algorithm is published in the most recent NIST FRVT Ongoing report (Nov. 19, 2019) under the submission `rankone-008`. The following table provides a summary of the results, as well as a comparison against the median performance for each metric. Note that FRR refers to the False Reject Rate and the FAR refers to the False Accept Rate.
Rank One is one of the few vendors achieving better than median performance across every single performance metric. This balance enables the Rank One algorithm to support nearly any face recognition application.
Of course, these improvements would not matter to Rank One’s integration partners and end-users if they were not readily available for use. Fortunately, unlike the antiquated practice in the biometrics industry of requiring an entirely new license purchase to receive algorithm improvements — a business practice that has resulted in critical national security systems using grossly outdated algorithms despite paying continued maintenance and support fees — Rank One practices Evergreen Licensing, which provides a transparent path to receive continual algorithm improvements.
Rank One’s CEO Brendan Klare had the following to say about this release: “This release represents the largest development effort ever undertaken by Rank One to produce a new algorithm. Through tremendous leadership within our R&D team, Rank One has established a new foundation that not only enabled the improvements delivered in this release, but has also primed us for a powerful v1.21 with expected delivery in March 2020. Perhaps our only regret is that we likely contributed to global warming through the trillions of compute cycles spent optimizing the v1.20 algorithm, though we will seek to offset this impact through our charitable donation program.”
Interested in the ROC SDK version 1.20? Contact our support team to upgrade your version of the SDK, or our amazing business development team to begin a new evaluation of the SDK.