ImageNet-E: Benchmarking Neural Network Robustness via Attribute Editing

Background Size Position & Direction ILSVRC2012_val_00005614 ILSVRC2012_val_00004577 ILSVRC2012_val_00009393 ImageNet-E Figure 1. Examples of the proposed ImageNet-E dataset. In contrast to adversarial examples or datasets like ImageNet-C [22] who add perturbation or corruptions to original images, we edit the object attributes with controls of backgrounds, sizes, positions and directions.