When you are using axis-aligned bounding boxes for training a model background, features are included with each rotated object, reducing the model’s ability to differentiate the objects of interest from the background imagery. The rotated boxes detected by ODTK (b) address this issue and better fit the outline of the objects.Īpplications that may depend on detection of rotated objects and features include remote sensing (Figure 1), text detection “in the wild,” medical physics, and industrial inspection. The axis-aligned ground truth boxes (a) overlap one another and each contains a mixture of classes (person and motorcycle). An example of ODTK detecting rotated boxes. (b) Rotated bounding-boxes detected by ODTK for the same imageįigure 3.(a) Axis-aligned bounding-box ground truth for COCO validation image.The addition of an angle parameter helps describe its location and outline with greater precision than an axis-aligned box. In the real world, some objects cannot be described as a simple rectangle and require even more parameters. Now you can describe the bounding box of an object using xmin, ymin, width, height and θ. An additional parameter is needed to reduce the difference between the area of the object and the bounding box that describes it the object angle relative to the vertical axis, θ (theta). Try the calculation yourself!įor rectangular objects, or any objects with a high aspect ratio (tall and thin, short and fat), the difference is even greater. The area of the bounding box is twice that of the square that you are attempting to describe. For example, try to describe a square that has been rotated by 45° using the four bounding box parameters.