Complementary Convolution Residual Networks for Semantic Segmentation in Street Scenes with Deep Gaussian CRF

Abstract

To understand surrounding scenes accurately, the semantic segmentation of images is vital in autonomous driving tasks, such as navigation, and route planning. Currently, convolutional neural networks (CNN) are widely employed in semantic segmentation to perform precise prediction in the dense pixel level. A recent trend in network design is the stacking of small convolution kernels. In this work, small convolution kernels (3 × 3) are decomposed into complementary convolution kernels (1 × 3 + 3 × 1, 3 × 1 + 1 × 3), the complementary small convolution kernels perform better in the classification and location tasks of semantic segmentation. Subsequently, a complementary convolution residual network (CCRN) is proposed to improve the speed and accuracy of semantic segmentation. To further locate the edge of objects precisely, A coupled Gaussian conditional random field (G-CRF) is utilized for CCRN post-processing. Proposal approach achieved 81.8% and 73.1% mean Intersection-over-Union (mIoU) on PASCAL VOC-2012 test set and Cityscapes test set, respectively.