Delving Deeper into Anti-Aliasing in ConvNets

Aliasing refers to the phenomenon that high frequency signals degenerate into completely different ones after sampling. It arises as a problem in context of deep learning downsampling layers are widely adopted architectures reduce parameters and computation. The standard solution is apply low-pass filter (e.g., Gaussian blur) before (Zhang in: ICML, 2020). However, it can be suboptimal same across entire content, feature maps vary both spatial locations channels. To tackle this, we propose an adaptive content-aware filtering layer, which predicts separate weights for each location channel group input maps. We investigate effectiveness generalization proposed method multiple tasks, including image classification, semantic segmentation, instance video image-to-image translation. Both qualitative quantitative results demonstrate our approach effectively adapts frequencies avoid aliasing while preserving useful information recognition. Code available at https://maureenzou.github.io/ddac/ .