Enlarging Effective Receptive Field of Convolutional Neural Networks for Better Semantic Segmentation

Recently, convolutional neural networks have shown powerful capability in different fields of computer vision, and have become the most effective means for dense prediction problems such as semantic segmentation. However, methods based on fully convolution network(FCN) are inherently limited to the size of the receptive field for each pixel, which leads to the bad performance of predicting object boundary. In this paper, we propose a novel deep neural network module, namely group dilated convolution(GDC), to effectively enlarge the receptive field, and a top-to-down pathway network is exploited simultaneously. The idea is that dilation convolution with different ratios can cover features of different scales, which shows a significant Mean IOU improvement in comparison with the baseline network.