Revisiting Sliced Wasserstein on Images: From Vectorization to Convolution

  • 2022-09-22 03:04:20
  • Khai Nguyen, Nhat Ho
  • 0

Abstract

The conventional sliced Wasserstein is defined between two probabilitymeasures that have realizations as vectors. When comparing two probabilitymeasures over images, practitioners first need to vectorize images and thenproject them to one-dimensional space by using matrix multiplication betweenthe sample matrix and the projection matrix. After that, the sliced Wassersteinis evaluated by averaging the two corresponding one-dimensional projectedprobability measures. However, this approach has two limitations. The firstlimitation is that the spatial structure of images is not captured efficientlyby the vectorization step; therefore, the later slicing process becomes harderto gather the discrepancy information. The second limitation is memoryinefficiency since each slicing direction is a vector that has the samedimension as the images. To address these limitations, we propose novel slicingmethods for sliced Wasserstein between probability measures over images thatare based on the convolution operators. We derive convolution slicedWasserstein (CSW) and its variants via incorporating stride, dilation, andnon-linear activation function into the convolution operators. We investigatethe metricity of CSW as well as its sample complexity, its computationalcomplexity, and its connection to conventional sliced Wasserstein distances.Finally, we demonstrate the favorable performance of CSW over the conventionalsliced Wasserstein in comparing probability measures over images and intraining deep generative modeling on images.

 

Quick Read (beta)

loading the full paper ...