Abstract: Fusion of the extracted high-order features to obtain a better representation by capturing the complex correlation between features has always been a research focus in visual tasks. As a simple and effective high-order feature interaction representation, bilinear representation has achieved remarkable results in many visual tasks: fine-grained image classification, semantic segmentation and so on. However, bilinear pooling has not been widely used due to the bilinear representation up to hundreds of thousands or even millions of dimensions. In this paper, we propose grouping bilinear pooling (GBP) that the representation captured by GBP can achieve the same effect with less than 4% parameters compare with full bilinear representation. This more compact representation largely overcomes the high redundancy of the full bilinear representation, it greatly reduces the computational cost of introducing bilinear pooling in visual tasks. The effectiveness of the proposed GBP is proved by experiments on the widely used fine-grained recognition datasets.