Help understanding bounding box coordinates

[agaeki]

Hello,

I'm trying to write a program that can recognise text from an image and highlight in the image where the text is, but the bounding boxes returned by engine.detect_words don't seem to have coordinates that make any sense in the image e.g. text that is on opposite sides of the image have similar coordinates, coordinates are sometimes out of range, height and width don't seem to be in pixel units. Can anyone help me understand how the coordinates of RotatedRect::center() work?

Thanks,
agaeki

[robertknight]

e.g. text that is on opposite sides of the image have similar coordinates, coordinates are sometimes out of range, height and width don't seem to be in pixel units.

The coordinates returned by eg. RotatedRect::center are in pixel coordinates of the input image. You can use the bounding_rect method to get an axis-aligned bounding rect. There is no guarantee of the orientation of the rect (ie. which way it's "up" vector points). You can use orient_towards to normalize this.

The rotated rects are formed by finding groups of connected text pixels in the input image and then determining the smallest rotated rect which encloses them. If the word is rotated, it is possible that some corners of such a rect may extend outside the boundaries of the input image. You can visualize the binary mask from which the rotated rects are formed by using the --text-mask option of the ocrs CLI tool:

# Generates text-mask.png
ocrs image.png --text-mask

You can visualize the word rects using the --png option of the ocrs CLI tool:

ocrs image.jpg --png --output annotated.png

The code which draws annotated.png is here.