search

akarshzingade / image-similarity-deep-ranking

369 stars 103 forks source link

Evaluation metric #3

Closed IAmAbdusKhan closed 6 years ago

IAmAbdusKhan commented 6 years ago

Hey Akarsh . Did you implement some evaluation metric also so as to quantify the results ?

akarshzingade commented 6 years ago

Notation: 1) Dqp - Distance between query image and positive image. 2) Dqn - Distance between query image and negative image. 3) g - Gap parameter.

The evaluation metric would be the hinge loss. The loss should ideally be zero, which would mean that Dqp is lesser than (Dqn - 'g'). In other words, the difference between Dqp and Dqn should be less than -g. Also, if the difference between Dqp and Dqn is less than 0 (this would mean the loss is less than 'g'), the model is doing good. So, the overall hinge loss is a good evaluation metric.