deformable convolutional networks (dcns)

These augment the spatial sampling locations in the modules of CNNs with additional offsets and learn the offsets from target tasks, capturing geometric transformations better.