Superb work! I stumbled over the same idea of using BAD in implementing PDL::Image2D::crop (see https://github.com/PDLPorters/pdl/commit/489cec5c6466ee513f5020d0d30a495474a35837) to achieve what in PDL is otherwise difficult/impossible: raggedly-sized ndarrays. (I also had to update the implementation of which to fill the rest of its output with -1 to use that as a badvalue).
Comment on Re^2: PDL: Looking for efficient way to extract sub-images, by finding bounding boxes of "objects" (BAD is good for you!)