The benchmarks below are for an A72 Cortex system on a chip (SOC).
- Facedetector depends on the input image but it usually requires around 200ms.
- Each recognition models requires about 500ms to resolve. There are multiple recognition models. (e.g. face, mask, etc.)
The way the SAFR Embedded SDK works is that each recognition model is invoked separately. For example, if you do face detection + face recognition, age detection, and mask detection then the SAFR Embedded SDK would require 200+500+500+500 = 1.7sec to finish if you run each item sequentially. However, you can run some model calls in parallel on multiple cores. Things like mask detection should be run before doing face recognition since you need to first determine if a mask is worn and then apply either the FaceRecognitionModelType_Regular or FaceRecognitionModelType_MaskOptimized, as appropriate.