Recognition results with "test only" worse than with ocr_predictor pipeline

[hanshupe007]

I am using the ocr_predictor function for text detection and recognition and now wanted to fine-tune the recognition model.

To create a test dataset, I used the same detection model as in ocr_predictor to crop all words. My expectation was that when I ran the same recognition model as before, using the --test_only parameter, it would give me exactly the same results as the ocr_predictor function.

However, for some reason, the results differ. It seems like ocr_predictor returns better results than when cropping the images first using the same detection model and then running the recognition model on those cropped images.

For example, in the image below, ocr_predictor correctly recognized "33," while running the models separately returns "|33."

What could be the reason for these different results? Is there any preprocessing happening inside ocr_predictor that is not part of the model inference?