For every recognized item we return a score representing similarity between the matched object from the query image and the most similar reference image from the item.
Scores are represented with values from 0 to 100. The more similar are the matched objects the higher the score. Scores can be also seen as a measure of confidence that the matched objects are indeed visually similar.
Which one is the best match?
The list of matches that the Image Recognition API returns is sorted based on their recognition scores from the best one to the weakest one. The best one is the first in the list.
How many matches does the Image Recognition API return?
Currently, we always return all matches that have a sufficiently high recognition score. The threshold controlling the scores has been set up experimentally to work well in most application scenarios. If your app needs to use only a limited number of the best matches you need to implement it yourself.
Examples and their scores
Below we show several examples of visual effects reflected by the recognition score. Note that if several of such conditions occur simultaneously then the corresponding effects on the score will be combined.
Example 1: Recognition scores reflect the degree of visual similarity between the query object and the matched reference images.
Example 2: Recognition of a blurred object results in a lower score.
Example 3: Recognizing a tilted object results in a lower score.
Example 4: Recognizing a small object hidden within a cluttered scene produces a lower score.