Amazon upgrades facial recognition for challenging images and opens Alexa to device developers
Amazon has updated the face detection, analysis, and recognition features of its Rekognition software to improve its ability to detect faces, increase its matching accuracy, and gain more information from faces in images. According to a company blog post, the enhancements are available at no extra cost, and do not require machine learning experience.
The updates allow Amazon Rekognition to detect 40 percent more faces from images with pose variations, occlusions, illumination variations, and low image quality. The rate of false detections is also reduced by 50 percent, which the company says allows its customers to deliver better user experiences for applications like automated profile photo review, according to the post by Rekognition project lead Ranju Das and Senior Product Manager Venkatesh Bagaria.
The “best” matches feature now returns 30 percent more correct images when searching against a large database, and the similarity scores of face matches are now more consistent for images with challenging lighting, pose, and appearance variations, which the post authors write allows customers to use higher confidence thresholds, avoid false matches, and reduce human reviews.
The post describes the challenges facial recognition systems have with faces which are looking away, or tilted in uncommon angles, such as upside down. High or low contrast lighting, shadows, motion blurring, and low-resolution images, such as those on identity documents, can all pose problems for face detection and recognition algorithms, it says.
As a protracted public debate over the use of the technology by law enforcement agencies continues, the blog post also notes that for use cases involving civil liberties or customer sentiment, best practices, a confidence level of at least 99 percent, and human review are recommended.
New SDK to bring voice interaction to more devices
Amazon has released its Alexa Mobile Accessory Kit to developers to integrate the voice assistant with a range of devices, and Qualcomm has made a reference design available for integrating Alexa with smart headsets, according to a separate blog post.
The SDK already supported a range of wireless Bluetooth headphones, including from Bose, Jabra, and Sony, and has now added support for smartwatches, portable Bluetooth speakers, and traditional headphones.
“For device makers, there is no need to develop and maintain a custom Alexa app. By leveraging the Amazon Alexa App that is readily available for Android and iOS, device makers accelerate product development, software integration, and testing time, while reducing product support costs,” explains Alfred Woo, a principal product manager for Alexa Voice Services. “After launch, their products are automatically updated to include new Alexa features and functionality when they arrive in the Amazon Alexa App.”
The Qualcomm reference design is available for $299, though The Verge cautions that it will likely be some time before new integrations reach commercially available products.
A pair of Canadian credit unions recently made their banking services available through Alexa, with transactions secured by voice biometrics.