Open atv2016 opened 2 years ago
They use different face detection mechanisms, and they make different mistakes
First, you can try is increasing det_prob_threshold
threshold. Second the option I described here:
https://github.com/exadel-inc/CompreFace/issues/966#issuecomment-1332528587
hi @pospielov , @atv2016 can you please explain how to do a custom build? the information in the docs is very abstract. Which google drive we need to upload the custom model to? We need an on-prem solution and google drive is not an option. Do you know how I can do it?
I've recently switched to using this custom build, in order to get more unknown detections. This seems to work, but one thing i noticed it doesn't seem to detect as much faces/from a far anymore.
They use the same face detection mechanism i assume, so is it discarding more faces up front? Is there anyway to adjust this? With the default build i could detect people and their faces from at the end of the road, with the new arcnet build you have to be less then 2 meters away it seems.