Multi-kernel non-local neural network for semantic segmentation

As a milestone in semantic segmentation, Non-Local Block (NLB) efficiently enhances the ability of regular convolutional neural networks in capturing long-range dependencies. From the view of mathematical modeling, NLB is based on a single Gaussian kernel. Existing works suggest that multi-kernel methods generally get more powerful performance in edge detection, which is crucial to image segmentation. Motivated by this consideration, we design a Multi-Kernel Non-local Block (MKNLB). As expected, the proposed MKNLB exhibits excellent behaviors when being used in semantic segmentation. Additionally, with the distributive law of matrix multiplication, the complexity of its implementation is comparable to that of the standard NLB. Theoretical analyses and preliminary experiments on benchmark datasets both support the same conclusions.