Availability of Quantized (INT8) Model for Latest MediaPipe Face Mesh (478 Landmarks)

[dhilshad]

I am looking to run Google's MediaPipe Face Mesh (478-point landmark model with iris tracking / Attention Mesh) efficiently on iMX Platforms. Does NXP currently provide an official quantized version (e.g., INT8) for the 478-point Face Landmarker model?

If not currently available, any Plans to release in future?

I have tried to quantize this original model from google using eIQ toolkit, but the accuracy is reducing drastically. When I quantize the Google's old Facemesh model (486 Landmarks) same way, i am able to get decent performance. Has NXP has the same observation?

Anyone else has tried quantising Google's latest FaceMesh(478 Landmark) model?