Version 1.19 of the ROC SDK is now available! This release includes the following improvements and additions to our industry leading face recognition software libraries:

  • Improved face recognition accuracy 
  • Improved facial clustering 
  • Added homomorphic encrypted matching
  • Added GPU accelerated enrollment
  • Added Facial hair estimation 
  • Improved video processing API’s

Accuracy Improvements

Rank One ships two face recognition algorithms: ROC_FR and ROC_FR_FAST. Our standard algorithm is ROCFR, which is always submitted to NIST FRVT Ongoing and delivers top-tier accuracy alongside industry leading efficiency. We also deliver a ROC_FR_FAST, which is an even more efficient algorithm that is tailored for embedded devices and high-throughput video processing. Both algorithms were significantly improved with this release. 

ROC_FR Improvements

The focal points of accuracy improvements in ROCFR were for highly unconstrained imagery, young and old persons, and low resolution faces. 

The v1.19 ROCFR algorithm was submitted to NIST FRVT Ongoing and is listed in the most recent report as “rankone-007”. The v1.18 ROCFR algorithms was previously submitted to FRVT Ongoing and is listed as “rankone-006”. Here is the performance comparison between these two algorithms, where FNMR is the False Non-Match Rate and FMR is the False Match Rate

table1 These improvements represent roughly a 20% reduction in error rate in both Visa and Wild imagery.

ROC_FR_FAST Improvements

Massive accuracy improvements were delivered to the ROC_FR_FAST fast algorithm. On unconstrained imagery the error rates were reduced by nearly 50%. 


The following is an efficiency comparison between the two algorithms:


The following is an accuracy comparison between the two algorithms on internal datasets:


Improved Facial Clustering

Rank One made major changes to our facial clustering algorithms, which results in far more accurate identity grouping. Rank One is uniquely capable of performing identity clustering due our leading efficiency in enrollment speed, template size, and comparison speed. With this new enhancement our users will be able to rapidly triage and organize facial imagery.

Homomorphic Encrypted Matching

As face recognition continues to support payment systems and access control, template security becomes increasingly important. To address these challenges Rank One now offers the option to perform homomorphic encrypted matching. Using a patent-pending method to improve the efficiency of the process, this optional feature from Rank One will allow integrators to encrypt templates, and perform matching on encrypted templates without ever needing to decrypt them. Thus, the contents of the template will remain secure throughout the storage and comparison process. 

GPU Enrollment 

While Rank One is known for enrollment speeds on a CPU that outperform GPU-based solutions, in this release we now support the ability to generate templates on a GPU, allowing our users to tap into more hardware resources. 

Facial Hair Estimation

Rank One offers a wide range of facial analytics algorithms, and now users will have the ability to detect the presence and thickness of beards and mustaches. 

Improved Video Processing APIs

Processing streaming video is one of the most challenging applications in face recognition. Fortunately this process can be made much easier through the well designed APIs (as well as computationally efficient algorithms). Rank One is continuously improving our video tracking and processing API’s, and this new release makes this process even easier for integrators. 

Next Up: ROC SDK Version 1.20

ROC SDK version 1.20 is already deep into development with a scheduled release date of the first week of October. This release is expected to deliver major accuracy enhancements, as well a further rebalancing of accuracy across race and gender.

Share This