Wasserstein Distortion with Intrinsic $\sigma$-Maps

Wasserstein distortion is a recently proposed family of distortion measures, controlled by a width parameter $\sigma$, that lifts fidelity and realism into a common framework. In previous implementations, calculating the Wasserstein distortion between two images relied on a companion saliency map or manual tuning to specify the width parameter $\sigma$ for each location in the image. We introduce a novel scheme for automatically generating an $\sigma$-map from the image itself.