Segmentation models - Evaluate and improve

To improve a model it’s good to know how evaluate it.

  1. So first evaluate your model.

  2. Then find out how to improve your model.

Evaluate segmentation models

  • Loss
    What’s good? You want the loss as low as possible. You want the training error to be slightly lower than test error.
    The loss curve can give you an insight on what you should do to improve your experiment, as you can find here.

  • Accuracy
    What’s good? For a perfect model the accuracy is 1.
    The accuracy measures how often the model gets the predictions right. Accuracy is not the best metric to evaluate performance in a segmentation problem, and you should also consider the value of the f1 score.

  • F1 score
    What’s good? If your model is perfect, the f1 score is 1.
    The F1 score is a way to evaluate the model’s predictions to consider when the dataset is imbalanced, like in the case of a segmentation problem.

Improve segmentation models

  • High training and validation error
    Your model is underfitting.
    Solution:

    • Run your model longer, that is, increase the number of epochs.

    • Collect more data.

    • Try another model.

  • High discrepancy between training and validation error
    Your model is overfitting.
    Solution:

    • Try collecting more data.

    • Make sure that the data in your training and validation set are similar.

Was this page helpful?
YesNo