Image size matters

[nhanerc]

I tried different image sizes on a dense text image like a dictionary page. Here is the output: 2048x2048 resolution:

1024x1024 resolution:

This is an expected behavior?

[nhanerc]

The image is from https://github.com/open-mmlab/mmocr/blob/main/demo/demo_densetext_det.jpg

[D641593]

Hi, the weights are all trained on the scene text detection datasets. If you want to apply it to dense text images, you need to retrain on it. I think this is because the size of the text in the image is too small and is seen as noise in the scene text detection dataset.