Entropy-Based Active Learning for Object Detection With Progressive Diversity Constraint
Created: =dateformat(this.file.ctime,"dd MMM yyyy, hh:mm a") | Modified: =dateformat(this.file.mtime,"dd MMM yyyy, hh:mm a")
Tags:
Annotations
image set ac-quired by the intra-class rejection process tends to favorcertain classes (i.e. the majority classes), leading to severeclass imbalance. * show annotation
- adaptively providing more budgets for theminority classes than the majority ones * show annotation
this addresses issue of classes that are not present in the image → i.e. no instance of this class?
==H(Ii,k) = −pi,k log pi,k −(1 −pi,k) log (1 −pi,k)== * show annotation
NMS cannot deal withthe instance-level redundancy, i.e. instances with similarappearances presenting in the same context, which is sup-posed to be reduced in active acquisition * show annotation
k-th instance * show annotation
pi,k is the confidence score predicted as the fore-ground of a certain category and 1 −pi,k as the background * show annotation
image-level basic detection entropy * show annotation
==H(Ii|DS) = ∑k∈[t] H(Ii,k)== * show annotation
Learn Loss [38] employs holistic image-level featuresfor uncertainty estimation and with the task-free loss predic-tion module, it directly evaluates how much information anunlabeled image contributes. * show annotation