Efficient Compact Bilinear Pooling via Kronecker Product

Bilinear pooling has achieved excellent performance in fine-grained recognition tasks. Nevertheless, high-dimensional bilinear features suffer from over-fitting and inefficiency. To alleviate these issues, compact (CBP) methods were developed to generate low-dimensional features. Although the existing CBP enable high efficiency subsequent classification, themselves are inefficient. Thus, inefficiency issue of is still unsolved. In this work, we propose an efficient method solve problem inherited thoroughly. It decomposes huge-scale projection matrix into a two-level Kronecker product several small-scale matrices. By exploiting ``vec trick'' tensor modal product, can obtain feature through decomposed matrices speedy manner. Systematic experiments on four public benchmarks using two backbones demonstrate effectiveness proposed recognition.