Added heuristic for bounding bbox ordering

[MLDovakin]

I have been using your opensource framework for quite some time and recently I needed to do a text detection task. The important part is that I had to keep the word order from left to right, but I saw in the documentation that there was no such feature.

I first tried using opencv bounding ordering but that didn't work because there were too many overlapping bounding bboxes. Then I tried to sort by height and set the Y threshold as the bounding bbox should differ minimally in height. This gave much better results than other methods and I would like to add such functionality here

Here are examples of using Y-threshold ordering

1. The result of get_sliced_prediction()
   

2, Draw ordering bboxes result

Plotting code

def bbox_sort(a, b, thresh):
    bbox_a = a
    bbox_b = b
    
    if abs(bbox_a[1] - bbox_b[1]) <= thresh: 
        return bbox_a[0] - bbox_b[0]
    
    return bbox_a[1] - bbox_b[1]

my_list = []

for ann in result.to_coco_annotations():
  ##type int so that there are no opencv errors when drawing lines

  current_bbox = ann['bbox']
  x = int(current_bbox[0])
  y = int(current_bbox[1])
  w = int(current_bbox[2])
  h = int(current_bbox[3])
  
  my_list.append((x, y, w, h))

thresh = 10
cnts = sorted(my_list, key=cmp_to_key(lambda a,b: bbox_sort(a, b, thresh)))

img = cv2.imread(f"/content/detect_images/output_01.jpg")
red = [0,0,255]

k = 0

font                   = cv2.FONT_HERSHEY_SIMPLEX
bottomLeftCornerOfText = (10,500)
fontScale              = 1
thickness              = 1
lineType               = 2

for i in cnts:
  q = cv2.circle(img, (i[0], i[1]), 5, red, -1)
  q = cv2.putText(q, f'{k}', (i[0],i[1]),font,1,(120,166,50),2)
  k += 1

cv2_imshow(q)

[fcakyon]

Can you please reformat your code and commit&push again as detailed in the contributing section of the readme :)

[fcakyon]

@MLDovakin thanks alot for your contributions!