currently object detection returns bounding boxes as absolute locations (pixel index), but the alternative (and used by tensorflow) is relative coordinates, as described here. Would it be possible to implement relative coordinates, perhaps selected via an optional arg? This would simplify compatibility with tensorflow post-processing scripts, is robust to resizing of images, and is preferable for tracking algorithms.
Hello @robmarkcole, this is a good suggestion, however the conversion from the absolute locations to the relative coordinates can be easily done on the user side.
A basic conversion in this case is simply,
(x_min/width, y_min/height, x_max/width, y_max/height)