0
$\begingroup$

Reducing images size will cause a loss of information for sure. If a have a model that perform better on resized images (50x50) than on original size images (224x224), what can I deduce ?

There is a lot of noise in the images data, the model is not enough complex to learn from all the data or something else ?

$\endgroup$
2
  • 1
    $\begingroup$ Are these models trained on the same data but resized? Same architecture, same everything? It could be to a lot of reasons, even randomness (if you don't have enough training samples). ONE possible reason: 224x224 images can create filters that tend to pay attention on background data instead of the main thing, 50x50 image can pay attention only to the foreground image. $\endgroup$
    – Tom
    Commented Sep 28, 2022 at 16:52
  • $\begingroup$ Yes on same data, architecture and everything. It's a strange behaviour anyway your observation makes sense. $\endgroup$ Commented Sep 29, 2022 at 9:43

1 Answer 1

1
$\begingroup$

Using bigger image does not mean better results. It increases computaional cost. As stated in comment section, it might be creating filters that focus on the background of the image. Also, if dataset is noisy you need to apply some filter, for instance Gaussian noise.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.