Dual Rank-1 Tensor Attention Module for Convolutional Neural Networks

Abstract

Channel-spatial attention mechanisms have been extensively investigated in computer vision. However, it is still a difficult problem that how to efficiently utilize global and local contextual information laid in a feature tensor to generate an accurate 3D attention map. This paper proposes a novel attention module for convolutional neural networks named Dual Rank-1 Tensor Attention Module, which can reach a good balance between global and local contextual information utilization for attention map generation. In our module, given a feature tensor, we sequentially generate two rank-1 3D tensor attention maps, i.e., the initial rank-1 tensor attention map containing global contextual information, and the complement rank-1 tensor attention map containing partial local contextual information. Then, we obtain a 3D tensor attention map based on the combination of these two rank-1 tensor attention maps for feature recalibration. Experimental results on ImageNet-1K and PASCAL VOC datasets demonstrate that the proposed module can achieve competitive performance compared with other state-of-the-art attention modules. The source code will be available at https://github.com/KevinBHLin/.