One SDK, many limbs: real-time face and body tracking on mobile

What it does TengineKit is a C++ computer vision SDK from OPEN AI LAB that packs face detection, 2D and 3D facial landmarks, iris tracking, upper-body detection, and YOLOv5 object detection into one library aimed at mobile and Linux devices. It promises a simple API and very low latency across a wide range of phone CPUs.

The interesting bit The SDK chases a rare trifecta on mobile: best performance, simplest API, and smallest package size, all while running everything from age estimation to iris localization locally. Its benchmark table is unusually candid, showing 250 fps on a flagship Kirin 980 but also 56 fps on a budget Qualcomm 450B.

Key highlights

• Face detection plus 2D/3D landmarks, face attributes (age, gender, smile, glasses), iris landmarks, upper-body detection, and YOLOv5 in one SDK
• Benchmarked at 4 ms on a Kirin 980 and 18 ms on a Qualcomm 450B for face detection plus landmarks
• Ships with a downloadable Android APK demo and Linux shared libraries
• Apache 2.0 licensed

Caveats

• Hand detection, hand landmarks, and the separate “body detect google” and “body landamrks” pipelines are marked “Real-time, not yet on Mobile,” so mobile builds currently focus on face and upper-body tracking
• The last documented update is March 2021, which makes the README promise of continued updates a bit of a time capsule
• README typos and inconsistent feature naming suggest the project has seen little recent polish

Verdict A solid shortcut for Android developers who need fast, local face tracking without building a pipeline from scratch. Those needing hand or full-body landmarks on mobile, or recent updates, should look elsewhere.