CODEWITHSUNDEEP
pythonopencvevaluationprecision-recallcbir
Viktoria
Viktoria

Reputation: 543

How to evaluate CBIR system without having a ground truth?

As my database I have a bunch of images of the same class - which means I do have only one class of images.

After implementing a CBIR (feature extraction by histograms and calculating distance by euclidean - very naive approach, nothing fancy), I get the 25 best matched results. And now I would like to evaluate them.

All evaluation metrics like mean average precision and precision recall need ground truth data, which I don't have. Right now I'm just taking one image and calculate the similarity.

How can I evaluate without having any ground truth data?

Upvotes: 0

Views: 263

Answers (1)

thelaw
thelaw

Reputation: 588

Late reply, but I am working on a similar project. You can't really evaluate your approach as you don't have ground truth. What you can do is to run it for various images and calculate the mean of how many images you think are relevant each time. Even better would be to use human judges instead of your own judgement since one could say it would be highly biased.

Upvotes: 0

Related Questions