Does recognition text score normalise between zero to one?

[balandongiv]

In mmocr '0.6.0' (I maybe wrong about the version since I dont have the old copy in hand due to recently upgrade to 0.6.1 without making a backup of the old mmocr) , I notice the score for every character is in range [0, 1] This apply for both SAR and RobustScanner.

However, in '0.6.1' , the score was cap around 0.029 for both the SAR OR RobustScanner. Is this a a raw logits score?

May I know whether there is code changes made from 0.6.0 to '0.6.1' that result in this new behavior?

The issue can reproduce via the following notebook. Should there is changes made, may I know how to get the score in range [0, 1]

Thanks in advance for the time taken entertaining this post.

[xinke-wang]

Hi, it seems a bug caused by this PR https://github.com/open-mmlab/mmocr/pull/1119 , we'll fix this soon.

[balandongiv]

Thanks for the feedback @xinke-wang .