1. 程式人生 > >論文閱讀-(ECCV 2018) Second-order Democratic Aggregation

論文閱讀-(ECCV 2018) Second-order Democratic Aggregation

本文是Tsung-Yu Lin大神所作(B-CNN一作),主要是探究了一種無序的池化方法\(\gamma\) -democratic aggregators,可以最小化干擾資訊或者對二階特徵的內容均等化。從另一個work line,對特徵聚合後,作matrix power normalization(Abbreviated as MPN)可以有效提升二階特徵的表達能力,MPN在aggregation時,隱含地均等化二階特徵。基於以上資訊,提出了\(\gamma\)-democratic aggregators, 整合了sum池化和democratic 池化。
主要是可以改進MPN在GPU上支援不友好情況,這點類似於

Is Second-order Information Helpful for Large-scale Visual Recognition?

Code

Network Structure

Orderless feature aggregation

Result

Reference