The evaluation metric is Mean Average Precision at 20 (MAP@20) on the test set. A good explanation of the metric can be found [here](http://fastml.com/what-you-wanted-to-know-about-mean-average-precision/).

In our case, there is only one true corresponding image. So for a description, suppose your algorithm finds the corresponding image in the list of top 20 images you return, and is ranked i-th. Then MAP@20 gives a score of:

$$score=\frac{1}{i}$$

For example, if it's ranked 1st you get a score of 1, if it's ranked 2nd you get a score of 0.5, if it's ranked 3rd you get a score of 0.333, if it's ranked 20th you get a score of 0.05, etc. If the corresponding image isn't in your list of top 20 images you get a score of 0. The overall score is calculated as the average score on the test set.

We divide the test set into two splits: a public split and a private split. Results on the Leaderboard are evaluated on the public split and the private split is used for grading.