Multi-Path Spatial Detail-Aware Network for Semantic Segmentation

In the existing semantic segmentation, small objects and slender strip objects that are easy to blend into the surrounding background are generally difficult to be accurately identified and segmented. We propose a Multi-Path Spatial Detail-Aware Network (MSDNet) for semantic segmentation of traffic scenes. Specifically, an interactive Dual-branch Parallel Complementary module (DPC) based on the Transformer-CNN framework is firstly proposed to enhance the discriminative representation of similar feature encodings between different objects. Then, a spatial detail assistant module (SDA) with attention mechanism is designed to help the network obtain more detailed spatial detail information. Moreover, a Cross-layer Feature Aggregation module (CFA) is constructed for further aligning and aggregating feature maps of different resolutions in adjacent two layers, which can improve the segmentation performance of small objects. Extensive experimental results show that the proposed method achieves the accuracy of 81.16% mean Intersection over Union (mIoU) on the challenging Cityscapes dataset.