Fixed rounding order of bbox values for yolo format

[Dipet]

Issue: #679

Fixed incorrect rounding order for yolo bbox format.

Example:

# Current master
x, y, width, height = (1200.0, 367.5, 158.0, 703.0)
x_min, y_min, x_max, y_max = (1122.0, 17.5, 1280.0, 720.5)

# PR results
x, y, width, height = (1200.0, 367.5, 158.0, 703.0)
x_min, y_min, x_max, y_max = (1122, 17, 1280, 720)

[creafz]

For all other bounding boxes formats, Albumentations returns augmented coordinates as floats. May we should also convert int coordinates to float before returning them to preserve the same output data type?

[Dipet]

@creafz You a little bit confused me :)
Yolo format is normalized float format, so after transforms we always get floats.
As you can see after conversion to int I call normalize_bbox function that returns floats.

[marcocaccin]

@creafz your suggestion can be a welcome addition, but it is out of scope for fixing this bug (and yes, yolo is float and normalised).

I think it's low on the priority list, but I do find myself having to query the data types of the albu inputs and restoring them afterwards - reason: albumentations transforms are performed inside a tf.numpy_function, which requires strict output types.
Alternatively, just use type annotations and specify that the output bboxes of albumentations will always be float64, that's fair as well