An existing JIVISION RK3588 smart-glasses project covered on-device face detection and an Android SDK customer-test delivery. Its records list face boxes, confidence scores, five facial keypoints, pose and session identifiers. This review explains those outputs without adding accuracy, inference-speed or production-readiness conclusions.
Recorded configuration and delivery status
The original project record lists the following configuration.
| Item | Recorded configuration |
|---|---|
| Platform | RK3588 Android |
| Application architecture | arm64-v8a |
| SDK packages | Core AAR + Face AAR |
| Image library | OpenCV 4.10.0 |
| Inference implementation | YuNet CPU |
This is the existing project's configuration. It does not establish compatibility with other systems or devices, and a processor's available accelerator is not evidence that the SDK uses that inference backend.
Distinguish the outputs
| Output | Information in the project record | Distinction to preserve |
|---|---|---|
| Box and confidence | Observed face region and score | A detection score is not an identity or a calibrated accuracy |
| Facial Keypoints | Five facial keypoints | Relate coordinates to input images, rotation and scaling |
| Pose | yaw/pitch/roll |
Check units, coordinate conventions and validity conditions |
| Direction class | Front, left, right and unknown | Do not force an unknown state into a definite direction |
| Session identifier | trackId |
Applies within the session, not as a persistent identity |
Verify box and landmark mappings after image cropping, rotation or display scaling. The presence of head-orientation output fields is insufficient evidence of any specified angular accuracy. Validate intended uses with samples and business requirements.
Detection and identity recognition
The documented delivery covers face detection and the listed output fields. Identity registration, feature extraction, identity-database matching and name or document recognition are outside that documented scope. Session identifiers are not persistent identities across sessions. Offline describes where processing runs; compatibility, performance and sustained operation require verification for each target configuration.
Verification after customer testing
The record describes a customer-test delivery, not production readiness. Subsequent integration needs output checks for the actual camera, formats, rotation, partial blocking of the face, illumination, concurrency and lifecycle. Define separate test conditions and acceptance records for performance, sustained operation or software-package deployment.
Sources: RK3588 smart-glasses offline detection project and JIVISION's original project publication record.
Record date: 2026-07-31.
Related services: Android, mobile and embedded vision SDKs and project enquiries.