1
$\begingroup$

I am working with Python, scikit-learn, keras and with 450x540 rgb images of front-faced watches (e.g. Watch_1, Watch_2).

My aim to run an autoencoder or a Siemese Neural Network to find the most similar watches among them. However, I am not sure if I will get better results by comparing pairs of images or triplets of images. As it is defined in this research paper, triplets of images consist of one target image, one image which is (more) similar to the target image and one image which is not (so) similar to it.

Can someone explain me in simple terms why using triplets of images will (necessarily) yield better results than using pair images, as research papers like the previous one claim?

$\endgroup$
3
  • $\begingroup$ towardsdatascience.com/… $\endgroup$
    – Aditya
    Feb 23, 2018 at 13:19
  • $\begingroup$ Excuse me but I do not get what this has to do with my question about triplets and pairs... $\endgroup$
    – Outcast
    Feb 23, 2018 at 14:44
  • $\begingroup$ How it's done otherwise $\endgroup$
    – Aditya
    Feb 23, 2018 at 14:55

1 Answer 1

1
$\begingroup$

Loss functions work best when there are clear definitions of correct and incorrect. If everything is correct, there is no signal for training. That is means the end of training. Loss functions need weighted errors in order to have something to minimize. In the case of image similarity, the weighted error is distance between dissimilar images.

Image triplets are more useful because they contain more signal for training a loss function compared to just image pairs.

$\endgroup$

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.