Channelized Axial Attention – considering Channel Relation within Spatial Attention for Semantic Segmentation

Spatial and channel attentions, modelling the semantic interdependencies in spatial dimensions respectively, have recently been widely used for segmentation. However, computing attentions separately sometimes causes errors, especially those difficult cases. In this paper, we propose Channelized Axial Attention (CAA) to seamlessly integrate attention into a single operation with negligible computation overhead. Specifically, break down dot-product of two parts insert relation between, allowing independently optimized on each location. We further develop grouped vectorization, which allows our model run very little memory consumption without slowing running speed. Comparative experiments conducted multiple benchmark datasets, including Cityscapes, PASCAL Context, COCO-Stuff, demonstrate that CAA outperforms many state-of-the-art segmentation models (including dual attention) all tested datasets.