Anti-Alias Downsampling
Anti-Alias Downsampling (AA) aims to improve the shift-equivariance of deep networks. Max-pooling is inherently composed of two operations. The first operation is to densely evaluate the max operator and second operation is naive subsampling. AA is proposed as a low-pass filter between them to achieve practical anti-aliasing in any existing strided layer such as strided convolution. The smoothing factor can be adjusted by changing the blur kernel filter size, where a larger filter size results in increased blur.