Compact bilinear pooling via kernelized random projection for fine-grained image categorization on low computational power devices

Abstract

[EN]Bilinear pooling is one of the most popular and effective methods for fine-grained image recognition. However, a major drawback of Bilinear pooling is the dimensionality of the resulting descriptors, which typically consist of several hundred thousand features. Even when generating the descriptor is tractable, its dimension makes any subsequent operations impractical and often results in huge computational and storage costs. We introduce a novel method to efficiently reduce the dimension of bilinear pooling descriptors by performing a Random Projection. Conveniently, this is achieved without ever computing the high-dimensional descriptor explicitly. Our experimental results show that our method outperforms existing compact bilinear pooling algorithms in most cases, while running faster on low computational power devices, where efficient extensions of bilinear pooling are most useful.