The torch dataset in this release is designed to use a standard torchvision transformation pipeline. This means the transformation pipelines operate on the image alone and do not account for the position of the object. Future releases will perform geometric transforms by using the mask to determine where the object is in the image.