Pyramid Context Contrast for Semantic Segmentation

Abstract

Semantic segmentation plays a critical role in image understanding. Recently, Fully Convolutional Network (FCN)-based models have made significant progress in semantic segmentation. However, achieving the full utilization of contextual information and recovery of lost spatial details remains a huge challenge. In this paper, we present a semantic segmentation model based on pyramid context contrast and a subpixel-aware dense decoder. We propose first using the pyramid context contrast to exploit the capability of contextual information by aggregating multi-scale foreground representations in different background regions via the pyramid context contrast module. Then, we add a subpixel-aware dense decoder architecture to reuse features extracted from different decoder levels by pixel shuffle, which can reasonably resolve resolution inconsistency between feature maps. Next, we refine the boundary by utilizing spatial visual information about low-level features via a boundary refinement branch with addition of auxiliary supervision. The presented model was evaluated using the PASCAL VOC 2012 semantic segmentation benchmark and achieved a performance of 86.9%, demonstrating that the proposed model achieves considerable improvement over most state-of-the-art models.

Keywords

• Semantic segmentation
• pyramid context contrast
• subpixel convolution
• dense decoder
• boundary refinement