论文阅读-(ECCV 2018) Second-order Democratic Aggregation

本文是Tsung-Yu Lin大神所作(B-CNN一作),主要是探究了一种无序的池化方法\(\gamma\) -democratic aggregators,可以最小化干扰信息或者对二阶特征的内容均等化。从另一个work line,对特征聚合后,作matrix power normalization(Abbreviated as MPN)可以有效提升二阶特征的表达能力,MPN在aggregation时,隐含地均等化二阶特征。基于以上信息,提出了\(\gamma\)-democratic aggregators, 整合了sum池化和democratic 池化。
主要是可以改进MPN在GPU上支持不友好情况,这点类似于 Is Second-order Information Helpful for Large-scale Visual Recognition?

Code

Network Structure

论文阅读-(ECCV 2018) Second-order Democratic Aggregation_第1张图片

Orderless feature aggregation

论文阅读-(ECCV 2018) Second-order Democratic Aggregation_第2张图片

Result

论文阅读-(ECCV 2018) Second-order Democratic Aggregation_第3张图片

Reference

  • Second-order Democratic Aggregation

你可能感兴趣的:(论文阅读-(ECCV 2018) Second-order Democratic Aggregation)