Large Kernel Spatial Pyramid Pooling For Semantic Segmentation

Spatial pyramid pooling is growing to become an important component in the network for semantic segmentation. However, it faces a dilemma of using larger kernels for better global context and computation cost. Recent architectures like ASPP have tried to solve this problem by using atrous convolution to keep large reception field while reducing the cost. However, atrous convolutions bring new problems like the "gridding effect", and the large gap between convolutional points make it hard to extract features of small or narrow objects. Inspired by the idea of stacking small filters to simulate large kernels, we propose a Large Kernel Spatial Pyramid Pooling to address both sufficient receptive field while maintaining efficiency. Our approach is evaluated on PASCAL VOC 2012 dataset and Road Extraction Challenge dataset, and achieved better results than competing architectures.