I'm working on object counting problem with deep learning object detection methods (specifically, yolo and faster rcnn). Is there any known method for uncertainty qunatification for object counting (i.e. predicted count $\pm$ x)?
Asked
Active
Viewed 15 times
0
-
It depends on the problem. Can you apply bootstrapping? – Ggjj11 Feb 19 '24 at 22:46
-
Are you looking for a [tag:prediction-interval], rather than a [tag:confidence-interval]? There is a difference. – Stephan Kolassa Feb 20 '24 at 06:54
-
That said, you can always fit models with various quantile losses to predict conditional quantiles (which will give you prediction intervals, not confidence intervals, but it sounds like this is what you actually want). – Stephan Kolassa Feb 20 '24 at 06:56
-
Is it reasonable to apply bootstrapping? For example if I have 300 training images, then I would create, say, 30 bootstrapped samples each having 300 images drawn from the 300 training images with replacement. Then I train models on each samples, then run inference on the test data. Combining the results will give me a prediction bound? – Andrew Lee Mar 20 '24 at 21:29