CODEWITHSUNDEEP
testingcomputer-visionclassificationevaluationobject-detection
user961627
user961627

Reputation: 12747

ROC curves/classification accuracy, what types of evaluation metric to use when?

Object detection accuracy, precision and recall - where precision and recall are plotted on ROC curves- these are the most common ways of visualizing and evaluating the results of image classification algorithms.

But there are two ways in which tests can be carried out.

  1. The classifier can be used to classify objects in images that are already cropped and containing only an object to be classified.
  2. Test images can also be a large scene within which the classifier should draw bounding boxes around potential objects to detect them.

I'm trying to figure out when to use which form of evaluation.

Can we use ROC curves to show performance when the classifier is classifying images in the first way of testing?

The problem I'm working on is multiclass classification, beyond binary classification.

Update

Am I correct in understanding that the first type of testing is called 'Classification' and the second is called 'Detection'?

Upvotes: 1

Views: 494

Answers (1)

old-ufo
old-ufo

Reputation: 2860

Second sometimes also called "Localization".

You have ROC curves, when you have some parameter to change - to draw it. Otherwise you simply have classification error and confusion matrix.

You can take an example from results of PASCAL VOC http://pascallin.ecs.soton.ac.uk/challenges/VOC/voc2012/results/index.html

Upvotes: 3

Related Questions