Question about YOLO format in convert_bbox_to_albumentations

[christian-cahig]

Hi, I was looking through this part of convert_bbox_to_albumentations :

albumentations/albumentations/augmentations/bbox_utils.py

Lines 232 to 244 in 95a048d

	elif source_format == "yolo":
	# https://github.com/pjreddie/darknet/blob/f6d861736038da22c9eb0739dca84003c5a5e275/scripts/voc_label.py#L12
	bbox, tail = bbox[:4], tuple(bbox[4:])
	_bbox = np.array(bbox[:4])
	if np.any((_bbox <= 0) | (_bbox > 1)):
	raise ValueError("In YOLO format all labels must be float and in range (0, 1]")
	x, y, width, height = denormalize_bbox(bbox, rows, cols)
	x_min = int(x - width / 2 + 1)
	x_max = int(x_min + width)
	y_min = int(y - height / 2 + 1)
	y_max = int(y_min + height)

Suppose the call x, y, width, height = denormalize_bbox(bbox, rows, cols) yields

• x = y = 4.0
• width = height = 3.0

for rows = cols = 5. This would give x_min a value of 3.0 and x_max a value of 6.0, i.e., x_max is greater than cols. The current implementation seems to shift the bbox 1 pixel to the right (and 1 pixel downwards).

With that said, would it be better to replace the definitions of x_min and x_max (and similarly for y_min and y_max) as follows?

x_min = max(int(x - width / 2), 0)
x_max = min(int(x + width / 2 + 1), cols)

Having the + 1 in the calculation of x_max instead of x_min enlarges the bbox by at most 2 pixels to the left and to the right but ensures that the object is still enclosed by the bbox. The max(..., 0) and min(..., cols) ensure that x_min and x_max are within acceptable values.

[christian-cahig]

With that said, would it be better to replace the definitions of x_min and x_max (and similarly for y_min and y_max) as follows?

x_min = max(int(x - width / 2), 0)
x_max = min(int(x + width / 2 + 1), cols)

Having the + 1 in the calculation of x_max instead of x_min enlarges the bbox by at most 2 pixels to the left and to the right but ensures that the object is still enclosed by the bbox. The max(..., 0) and min(..., cols) ensure that x_min and x_max are within acceptable values.

There's a flaw in this idea: the bounding boxes will only get larger for every transformation. This might not be a good idea especially when using a lot of transformations.

[Dimfred]

Why do the denormalize and the normalize in the first step?
yolo => albumentations can just be transformed by:

x, y, w, h = bbox # from yolo
w_half, h_half = w / 2, h / 2

x_min, x_max = x - w_half, x + w_half
y_min, y_max = y - h_half, y + h_half

return

that omits the pixel errors in the first place. Is there a particular reason why it is not like this?

[Dipet]

It looks like a bug that needs to be fixed.

[Dipet]

Should be fixed by #924