This PR adds a small extension to the RotationTransform transform for object detection; the user can choose to rotate bounding boxes by approximating the underlying object as an ellipsis rather than a rectangle. This solves a common problem with using rotations, where the bounding boxes are enlarged (this can significantly hurt performance).
In the below image, the original bounding box is shown in red, the "normal" way of rotating the bounding box (current behaviour) is shown in blue. The elliptical method added by this PR is shown in green. As you can see, the rectangular boxes are significantly larger than the objects which the boxes are for. Source code for the below image
This PR adds this feature by overloading RotationTransform.apply_box and adding a new parameter box_method. The default behaviour is unchanged, but if the user sets box_method='ellipsis' they will get the tighter elliptical boxes. This also passes the linter :)
This PR adds a small extension to the
RotationTransformtransform for object detection; the user can choose to rotate bounding boxes by approximating the underlying object as an ellipsis rather than a rectangle. This solves a common problem with using rotations, where the bounding boxes are enlarged (this can significantly hurt performance).In the below image, the original bounding box is shown in red, the "normal" way of rotating the bounding box (current behaviour) is shown in blue. The elliptical method added by this PR is shown in green. As you can see, the rectangular boxes are significantly larger than the objects which the boxes are for.
Source code for the below image
This PR adds this feature by overloading
RotationTransform.apply_boxand adding a new parameterbox_method. The default behaviour is unchanged, but if the user setsbox_method='ellipsis'they will get the tighter elliptical boxes. This also passes the linter :)For reference: [Paper showing significant performance improvements] [Same feature in albumentations]