[BUG: Output] Incorrect word-level bboxes in >=v0.15.0

[mzilinec]

📝 Describe the Output Issue

Hi guys, I have just observed that in versions >=0.15.0, word-level (and character-level) bounding boxes seem to be way less accurate (or maybe scaled incorrectly?). This seems to happen across different documents.

I tried the minimal code below, which produces correct bboxes on 0.14.7 but fails on 0.16.3 - I’m sending some examples of the result.

#foundation_predictor = FoundationPredictor()
recognition_predictor = RecognitionPredictor()#foundation_predictor)
detection_predictor = DetectionPredictor()

predictions = recognition_predictor([images[0]], det_predictor=detection_predictor, return_words=True)

drawn_im = images[0].copy()
draw = ImageDraw.Draw(drawn_im)

def random_color():
    return randint(0, 255), randint(0, 255), randint(0, 255)

for i, cell in enumerate(predictions[0].text_lines):
    draw.rectangle(cell.bbox, outline=random_color())
    # print(cell.original_text_good)  # False
    if cell.chars:
       for word in cell.chars:
           poly = [(int(x), int(y)) for x, y in word.polygon]
           draw.polygon(poly, outline=(128, 0, 0))

[tarun-menta]

Hi @mzilinec. Thank you for the error report!

We've identified the issue, which is a regression from our latest model. We're working on updating the model ASAP to rectify. In the meantime, please use the older version of surya. I'll keep you updated when we release a fix.

[niz-ib]

same issue with v0.17.0