What's New
-
Faster extraction of biometric templates and reduced model weights. A lightweight version is now available for each of the up-to-date recognition methods. While minimizing accuracy degradation, lightweight models provide increase in inference speed ( ~40-50% for mobile CPU and ~20-25% for desktop CPU) an ~60% reduction in model weights used in template extraction.
-
Improved accuracy of Liveness Detection methods. We continue to improve the accuracy of our Liveness Detection algorithms. In this release you'll see a 40% increase in attack detection accuracy (decrease APCER@BPCER=0.05 on all attack types).
-
Redesigned TemplateIndex - functional and user-friendly. We redesigned the module responsible for storing the template database and 1:N search. Now you can:
- Dynamically (without recreating TemplateIndex) add and remove templates from the database.
- Name each template with an arbitrary alphanumeric ID (previously you had to store this information separately).
- In case of highly loaded applications, an asynchronous version of the new TemplateIndex can be used, which will allow simultaneous addition/removal of elements and 1:N searches.
-
Testing methods for evaluating SDK quality. This is a detailed guide, including the main tasks addressed by Face SDK, the quality metrics used and a set of scripts for testing. With this guide you'll be able to quickly test the core technologies of Face SDK for your specific use case and see how well it meets your needs.
-
Support for accelerated inference on Rockchip NPU. Support for NPU inference from the Rockchip manufacturer has been added for the Face detection and Face Template Extractor modules. Using the new inference will significantly speed up calculations even on the company's low performance devices.
Bug Fixes and Improvements
-
Filtering detection results by bbox size. Ability to filter detection bboxes by width and height. Ideal for filtering out small faces in the background.
-
Combining multiple detectors into a cascade to improve accuracy in complex cases. When you need to detect faces in the same entrypoint on different domain data, such as selfies, ACS camera photos and ID card photos, it can be difficult to ensure high recognition quality. Combine different face detectors into a cascade so you don't miss a single face!
-
C# Face SDK API update with support for .NET 8. Due to the expiration of the LTS for .NET 6, we have migrated our C# Face SDK API to .NET 8.
-
Fixed a bug in the SetBytes method of the Context class in Java and the Kotlin API that led to segfault on 32-bit devices.