Has anyone tried PTQ (Post training Quantization Fully INT8) of FaceMesh v2 model(478 Landmark points)?
We see significant accuracy degradation when we quantize this v2 model.
But the old FaceMesh model with 468 Landmark points give a better result even after INT8 Quantization.
Anyone else has similar observation?
Any Suggestions to improve the Accuracy of 478 model after PTQ?
If this model is not fit for PTQ, is there any other techniques to quantize this model?