Correlation between number of Landmarks and inference latency

[aris-konstantinidis]

Dear @PaulTR

Again, thanks for the awesome work!

In our project we only need 12 (main body joints) out of the 33 inferred landmarks. Would a model predicting only 1/3 of the landmarks significantly reduce inference latency? In that case, could the existing pose_landmarker model be adjusted, or would a complete retraining be necessary?

Any input would be much appreciated.

Best regards, Aris