Is it just a question of training data parity?<p>Or are the current algorithms just failing on darker faces because of some inherent issue? shadows/lighting?edge detection?<p>It doesn't seem that hard to make a diverse dataset, so why are most of the models on the market awful at it?
I think lighting is the most important factor here.<p>It certainly can be an issue with training data or a matter of feature selection. On the other hand detecting features on dark skin is simply a lot harder in most cases. I did skin analysis for diagnostic support for dermatologists and feature extraction is simply more difficult. You often need much tighter limits to separate areas.<p>This was with good cameras in an environment with stable lighting conditions. Cameras for surveillance plainly lack the dynamic range of different light levels to provide enough contrast for analysis and depending on the algorithm precision is very important.