What are the output of detection_predictor?

[yumikim381]

Hello! I was wondering what exactly detection_predictor returns?

Let's say output is the return value from DetectionPredictor([npimg]).
I get that when you do output [0]["words"], you get an array with 4 coordinates of the bounding boxes and score,
but the coordinates seems to be all starting with "0. .. " .Like this :
[[0.6015625 0.9375 0.70898438 0.97265625 0.58779669] [0.484375 0.93847656 0.55566406 0.97265625 0.5817132 ] [0.22265625 0.93652344 0.29394531 0.97558594 0.56156856] [0.13085938 0.93847656 0.19238281 0.96777344 0.60816777]
My question is how are the coordinates ordered: i m guessing they are xmin, ymin, xmax, ymax but is the order correct?
Also why are the coordinates so small? How can i get the coordinate of bounding boxes in the original image?
Thank you and have a nice day!

kind regards,
Yumi

[felixdittrich92]

Hi @yumikim381 👋,

Correct the format is [xmin, ymin, xmax, ymax, score] and the coordinates a relative so you need to multiply the x values with the width and y values with the height of your image :)

[yumikim381]

That helps so much!
Thanks a lot for your swift reply

[yumikim381]

One more question, if i give such images of one line to DocTR detection_predictor, i get no detection results.

Do you by chance know if this phenomen is due to the lack of words in those images or due to preprocessing(scaling)?
I know that docTR line detector is not trained to detect on images with single line on it, but doctr's ocr_predictor seem to be able to detect on single line images.

Above is the result of doing :

result = predictor(doc)
result.show()
synthetic_pages = result.synthesize()
plt.imshow(synthetic_pages[0]); plt.axis('off');

So i m wondering if this is bug in my code. . Do you know how to fix this?

[felixdittrich92]

CC @odulcy-mindee @frgfm i think we should move this into the detection predictor call !? wdyt ?

[yumikim381]

Ohh is this for removing paddings from detection results before giving it to recognizer? Do you know why it is necessary to do this?

[felixdittrich92]

Each page is resized to 1024x1024 (default value which was also used for training) for detection ..additional there are multiple options like symmetric_pad and preserve_aspect_ratio so with remove_padding we keep all these option values into account to transform the results back to the original input page shape :)

[yumikim381]

Ahh okay thank you so much for the infos!!

[felixdittrich92]

Already opened a PR to fix this: #1627

[appsembly-team]

I am getting like this. Can someone help related to this to xmin, ymin, xmax, ymax and score? Thanks