A Graph-Involved Lightweight Semantic Segmentation Network

To extract cues for pixelwise segmentation in an efficient way, this paper proposes a lightweight model that involves graph structure in the convolutional network. First, a cross-layer module is designed to adaptively aggregate hierarchical features according to the feature relations within multi-scale receptive fields. Second, a graph-involved head is presented, capturing long-range channel and feature dependencies in two sub-domains. Specifically, channel dependency is acquired in a compact spatial domain for context-aware information, while the feature dependency is obtained in the graph feature domain for category-aware representation. Afterwards, by fusing the features with long-range dependencies, the network outputs the segmentation results after a learning-free upsampling layer. Experimental results present that this model remains light while achieving competitive performances in segmentation, proving the effectiveness and efficiency of the proposed sub-modules. ( https://github.com/xia-xx-cv/Graph-Lightweight-SemSeg/ ).