Headpose vectors scaling by 1.5

[saeidn95]

I wonder where the factor 1.5 (

human/src/face/angles.ts

Line 97 in 6239631

const size = Math.max(face.boxRaw[2] * imageSize[0], face.boxRaw[3] * imageSize[1]) / 1.5;

) in the size of the x and y coordinates come from.

[vladmandic]

honestly, don't remember exactly, been a while, but i believe its because box size is ~1.5x mesh size as a general item (although it can vary slightly).

[vladmandic]

it's not, that's bug in a facemesh model itself.
I've reported it to mediapipe long time ago, but they didn't do anything about it.

[saeidn95]

What do you mean by "It's not" Do you mean that there is bug in the network causing z not to be in the same scale as x? and to fix that your solution is to scale z by 128 separate from x and y that you scale to box size?

[vladmandic]

yes, that's what I mean.

[saeidn95]

Very good! Mediapipe has a new version that might have overcome this bug. But it seems they no longer provide the model in the tflite format.

[saeidn95]

Here is my final understanding of how you account for the scaling bug in mediapipe model. Tracing through the whole code from the mesh detection all the way to calculateFaceAngle in line (

human/src/face/angles.ts

Line 101 in 6239631

const yAxis = normalize(subVectors(pts[1] as Vector, pts[0] as Vector));

), the whole sequence of operations reduces to scaling x and y by a factor of ~1.5 to match the scale of z value; or alternatively scale z by (1/1.5) to match the scale of x, and y.