Realtime Global Attention Network for Semantic Segmentation

TL;DR

This paper introduces RGANet, a real-time global attention neural network for semantic segmentation that uses depth-wise convolution and affine transformations, achieving state-of-the-art performance with an improved evaluation metric.

Contribution

The paper presents a novel global attention module integrated into a hierarchy architecture for real-time semantic segmentation, along with a new evaluation metric MGRID.

Findings

RGANet achieves leading performance on semantic segmentation benchmarks.

The global attention module maintains high inference speed.

MGRID improves evaluation accuracy for scattered ground-truth areas.

Abstract

In this paper, we proposed an end-to-end realtime global attention neural network (RGANet) for the challenging task of semantic segmentation. Different from the encoding strategy deployed by self-attention paradigms, the proposed global attention module encodes global attention via depth-wise convolution and affine transformations. The integration of these global attention modules into a hierarchy architecture maintains high inferential performance. In addition, an improved evaluation metric, namely MGRID, is proposed to alleviate the negative effect of non-convex, widely scattered ground-truth areas. Results from extensive experiments on state-of-the-art architectures for semantic segmentation manifest the leading performance of proposed approaches for robotic monocular visual perception.