Dlib + opencv webworker performance

Hi guys,

I hope somebody can suggest some solutions or the root of the issue.
I found this repo https://github.com/seanchas116/wasm-face-detection.
Basically dlib/opencv emscripten build.
I took it, removed all logic with canvas rendering and use only Face Tracking part, and it works fine.
The issue I'm facing with is if I run the build in a dedicated worker.
Performance drastically decreases. If I run it in the main thread I get ~60ms for face Track. If the same code is moved to a web worker performance drops down to ~180 -200ms per cycle (I measure only tracking function, all message potsings, data convertioans etc. are not included in execution time measurement).
How can I overcome the issue? I tried to change optimization options (-O3 by default), but got only execution time increasing.
Are there any specifics how wasms run in workers?