What is the confidence reported by readtext? How to get detector confidence?

[uoefeb]

Hi all,

When readtext() reports "confidence", what is the "confidence" it is reporting?

I am doing single character detection and recognition, and am trying to measure performance in terms of mAP. In order to do so, I'd like both the bbox detection confidence and the recognition confidence (i.e. $p(bbox)$ and $p(character|bbox)$ ).

How could I get these?

Thanks!

[uoefeb]

Based on #959, #538, and looking over recognition.py, it looks like the confidence score that is reported is an "empirical" geometric mean of the probability of the character sequence, and has nothing to do with the bbox, is this right?

Does anyone know if the recognizer has the ability to predict an empty string? (would that just be a bunch of $\epsilon$ or blank character predictions as discussed in the Shi et al. paper?

[uoefeb]

Does anyone know if the recognizer has the ability to predict an empty string?

Based on experimentation, the answer appears to be "yes." Between the detector and the recognizer in easyocr.readtext() I tried overwriting the image with both black and white as follows

        img_cv_grey = np.zeros_like(img_cv_grey) #overwrite black
        img_cv_grey += 255 # overwrite white

Black tends to predict empty strings or the number "3" with a very low confidence for some reason 🤷 .
White yields the empty string with a confidence of 0.0

---

Unfortunately, for my purposes, I'm not sure how useful that is in terms of a traditional object detector confidence score