Corrupted analysis, e.g. face lost, video too dark


#1

Hi,

We understand that the Affectiva scores are % level of confidence for the given metric.
It works nicely, but becomes problematic if the video is not fit for analysis for example no face shown…

What are the methods for optimising scores for these cases?
Is there a clear indicator for inadequate reading?
Should we use a combination of other metrics, indicating no face or low light etc.? What would be these indicators?

Thank you for your help
(C++, ubuntu, 4.0)


#2

Hi Richard, please refer to this page for getting more insights about getting optimal results. Yes you are correct you can use the faceQuality API to get an overview about the light.


#3

We have found the following indicators in relevancy:
onFaceFound
onFaceLost
brightness
glasses - may affect results
Anything else we might have missed?


#4

onFaceLost and onFaceFound are the callbacks on ImageListener API that give an indication when the SDK finds or losses a face. These callbacks happen first and if the SDK finds a face then the classifiers viz brightness, glasses, emotions, expressions etc are called. So if you want to discard the frames based on whether the face is found or not that you should focus on just the callbacks.