Multi-path segmentation network

We present a new type of convolutional network for semantic segmentation here. We tested it on several benchmark datasets, including PASCAL VOC, PASCAL Context and Cityscapes. It achieved superior performance compared to state-of-the-art segmentation methods. To increase segmentation accuracy, we design a special structure with multiple columns. The special structure creates much more paths for information flow. Therefore, it has the potential for more accurate segmentation. We propose the idea of multi-path design here, and hope it can help inspire new ideas.