消除Aliasing!加州大学&英伟达提出深度学习下采样新思路:自适应低通滤波器层
极市导读
Abstract
Introduction
提出一种新颖的自适应低通滤波器层用于解决Aliasing问题; 提出一种新颖的度量准则用于评价语义/实例分割的平移一致性; 通过ImageNet分类、Cityscapes分割以及COCO分割等任务验证了所提方案的有效性; 通过定量与定性实验表明了所提方法的可解释性。
Method
Spatial adaptive anti-aliasing
Channel-grouped adaptive anti-aliasing
Learning to predict filters
Analyzing the predicted filters
Experiments
Conclusion
CARAFE:Content-Aware ReAssembly of FEatures LPF:Making Convolutional Networks Shift-invariant Again. WeightNet Revisiting the Design Space of Weight Networks
赞 (0)