DAABNet: depth-wise asymmetric attention bottleneck for real-time semantic segmentation

With the increasing demand for the real-world applications such as autonomous driving and video surveillance, lightweight semantic segmentation methods achieving good trade-offs in terms of parameter size, speed and accuracy have attracted more and more attention. In this context, we propose a novel real-time semantic segmentation model. First, we design a two-branch depth-wise asymmetric attention bottleneck (DAAB) based on residual network to reduce the number of parameters and improve the inference speed. Particularly, an attention refinement module (ARM) is added in the DAAB module to make the information extracted from the two branches complement each other. Second, we design a strip pooling attention (SPA) module which combines the strip pooling module and the attention mechanism to pay more attention to strip-shaped objects and to establish long-range dependencies between discrete distributed regions, so that to address the problem of poor segmentation of strip shape objects. In addition, we also fuse information from different stages to compensate for the loss of spatial information, thus improving the ability of the network to segment small objects. Experiments on CityScapes and CamVid dataset demonstrate that the proposed method achieves impressive trade-offs in terms of parameter size, speed and accuracy. Code is available at: https://github.com/mhhz/DAABnet1.