A doubt regarding estimating the confidence over class prediction

[skbhat]

Hi,

Thanks for a very insightful paper titled "Estimating and Evaluating Regression Predictive Uncertainty in Deep Object Detectors".

I am planning to use it for assess the reliability of the prediction. I have a doubt regarding the inference stage. Suppose on a test image I run the predictive inference. Assume that I obtain a box which is classified as 'car' with a categorical probability .9. In your paper I could not see any direct way of finding how sure the model on this classification. Variance is estimated only for ground box co-ordinates. Do you compute the confidence on this prediction based on the determinant of covariance matrix of predicted bounding box?

Thanks for your time.