Open misterpeddy opened 4 years ago
Another idea on improving performance could be introducing a threshold on displacements from the face before running the inference model?
I think the idea of frame-by-frame could be good, it may be better suited if we were to predict a user is about to touch their face by tracking the route?
I think the idea of frame-by-frame could be good, it may be better suited if we were to predict a user is about to touch their face by tracking the route?
Yeah, using a bit of physics to predict that could be a good idea, but I can see some (potential) caveats to that:
Currently the performance of the application is borderline -- if we do a bit more work per inference, we'll definitely start to slow down some more limited devices.
First we have to profile to understand where the cycles are going. How much compute resources is each of handpose, facemesh, inference, drawing utilities, and other utilities consuming?
To make things slightly more complicated, we have a few backends that all the models could be executing on: CPU, WASM, WebGL - Read more about it here
Ideas on how to improve performance: