Integrated crossing pooling of representation learning for Vision Transformer

In recent years, transformer technology such as ViT, has been widely developed in the field of computer vision. In the ViT model, a learnable class token parameter is added to the head of the token sequence. The output of the class token through the whole transformer encoder is looked as the final representation vector, which is then passed through a multi-layer perception (MLP) network to get the classification prediction. The class token can be seen as an information aggregation of all other tokens. But we consider that the global pooling of tokens can aggregate information more effective and intuitive. In the paper, we propose a new pooling method, called cross pooling, to replace class token to obtain representation vector of the input image, which can extract better features and effectively improve model performance without increasing the computational cost. Through extensive experiments, we demonstrate that cross pooling methods achieve significant improvement over the original class token and existing global pooling methods such as average pooling or maximum pooling.