Bounding boxes: VOC to darknet transform - why minus 1?

[umbrellait-maksim-titarenko]

Hello!

Exploring the issue of converting annotations for custom model training, I discovered that we apply the following transform to the VOC (see: scripts/voc_label.py):

    x = (box[0] + box[1])/2.0 - 1
    y = (box[2] + box[3])/2.0 - 1

where:

• bbox[:4] are "bndbox"-values: xmin, xmax, ymin, ymax - parsed from VOC xml-files,
• x, y assumed to be a coordinate of the center of "bndbox".

I'm a bit confused with that since:
e.g. we have a box with xmin, xmax = 10, 20
in this case, intuitively, the x-center should be at the point = 15
however, the code above will give us the value 14.0

My question is:

• Why do we apply those "minus 1"?
  Do the VOC-annotation coordinates differ from the generally accepted ones (0,0 - for the left top corner)?

[umbrellait-maksim-titarenko]

I see there was the commit to apply the "minus 1" operations:

@AlexeyAB, can you please comment on why was this done?

[Sarah20187]

same question here， Thanks！