Evaluating computer vision models

Vision metrics and data leakage between images.

20 minutesComputer vision

What you will learn

  • Avoiding leakage in image data
  • Analysing the model's errors visually

For classification: accuracy, the confusion matrix and the accuracy of each class. Inspect the errors with the images themselves: what confused the model?

Beware: images of the same person or the same patient in both training and testing inflate the accuracy falsely; split by person or patient, not by image.

Test under real usage conditions (lighting, camera, background).

Implementation steps with the tool

  1. Teachable Machine: Under the hood → the accuracy of each class and the confusion matrix.
  2. Orange: Confusion Matrix, then Image Viewer for the misclassified examples.

Quick review question

Why do we split X-ray images by patient?

Tools

Try it in the lab