An Autonomous Selection Method for Spatial Pooling Based on Chi-Square Test

Abstract

Spatial Pyramids is one of the most widely used models in image scene classification tasks. Although the default spatial pooling layout in the model is fine, some manually defined spatial pooling layout always outperforms them on many problems. More and more evidence shows that spatial pooling layout is Task-Dependent, that is to say, different spatial pooling layouts are optimal for different datasets. For this reason, we propose a method based on Chi-Square(χ ² ) Test to autonomously select the optimal spatial pooling layout on a pooling layout candidate set. The proposed method especially works on the representation level, which is magnitude faster than selection by classification performance. Experimental results show that pooling layouts selected by our method have a better classification performance than the default spatial pooling layout and some other existing methods.