Semantic Segmentation in Autonomous Driving using Multi-scale Feature Fusion and Graph Network

Abstract

Semantic segmentation in urban scenes is an important task in computer vision. However, there are many challenges in urban road scenes, such as category imbalance and complex backgrounds. These problems lead to unclear edge segmentation and fuzzy categorization of occluded objects in existing semantic segmentation methods in urban scenes, which limits the accuracy and robustness in practical applications. In this paper, we propose a model that recursively enhances edge feature representation combined with local spatial context. We address the problem of unclear edge segmentation by introducing Multi-scale Central Difference Convolution (MS-CDC) to fuse multi-scale edge features. The FeedBack Connection (FBC) module based on feature pyramid fuses the multi-scale features while recursively augmenting the original network to make the occluded objects more robust. Meanwhile, we design a Local Feature Extraction (LFE) module to capture pixel-wise relationships by constructing local pixel graph and center pixel graph. It can learn local contextual information to extract finer pixel features. Experimental results on the Cityscapes and Mapillary Vista dataset validate the effectiveness of our proposed model. Our model can achieve new results by 80.67 \(%\) and 45.5$%$ mIoU on the Cityscapes and Mapillary Vista val set. We open-source our code at \href{https://github.com/sanmanaa/segmentation-autodriving-graph-centralconv}{https://github.com/sanmanaa/segmentation-autodriving-graph-centralconv}