Abstract Semantic segmentation networks usually utilize special pyramid structure after encoder or combine low-level and high-level feature maps in decoder to capture multi-scale context information, which we term them feature combination and feature connection, respectively. However, both such frameworks are less valid with the fixed geometric structure or unsuitable interim. In this paper, we advocate Dynamic Attention Network (DAN) to solve these problems. First, we design a Deformable Attention Pyramid (DAP) module to perform a self-adjustable descriptor of high-level output, which utilizes deformable function to model geometric transformation. With DAP, semantic information can be captured effectively. Second, we propose a Fusing Attention Interim (FAI) module to guide the back-propagation of long-short range information in each level of decoder. We evaluate DAN on the challenging PASCAL VOC 2012 and Cityscapes segmentation benchmarks and find that it achieves state-of-the-art results without post-processing. Our observation can be concluded that the flexible structure that possesses dynamic attention mechanism is beneficial to learn multi-scale context information.
Paper
Full text
Dynamic attention network for semantic segmentation
Semantic Scholar · Computer Science · 2020
Abstract
Abstract Semantic segmentation networks usually utilize special pyramid structure after encoder or combine low-level and high-level feature maps in decoder to capture multi-scale context information, which we term them feature combination and feature connection, respectively. However, both such frameworks are less valid with the fixed geometric structure or unsuitable interim. In this paper, we advocate Dynamic Attention Network (DAN) to solve these problems. First, we design a Deformable Attention Pyramid (DAP) module to perform a self-adjustable descriptor of high-level output, which utilizes deformable function to model geometric transformation. With DAP, semantic information can be captured effectively. Second, we propose a Fusing Attention Interim (FAI) module to guide the back-propagation of long-short range information in each level of decoder. We evaluate DAN on the challenging PASCAL VOC 2012 and Cityscapes segmentation benchmarks and find that it achieves state-of-the-art results without post-processing. Our observation can be concluded that the flexible structure that possesses dynamic attention mechanism is beneficial to learn multi-scale context information.