SDBNet: Lightweight Real-time Semantic Segmentation Using Short-term Dense Bottleneck

A popular choice when designing a semantic segmentation model is to adopt a pre-trained Deep Convolution Neural Network (DCNN) as a backbone and add extra modules for better semantic representation and competitive segmentation results. However, the large number of parameters and substantial memory footprint of these DCNN architectures make these large models unsuitable for real-time applications on mobile devices. To address the issue, this study proposes a very lightweight model, called Short-term Dense Bottleneck Network (SDBNet). By staging a series of bottleneck blocks, an efficient module, termed SDB, is carefully designed and it provides diverse field-of-views for better contextualization of varied geometrical objects in a complex scene. For precise localization, a shallow branch is deployed in parallel to SDB which shares the spatial details with the SDB module at multiple stages. At the decoder end, a simple, yet effective feature refinement and semantic aggregation module is deployed for better context assimilation and region identification. The proposed model is evaluated using three public benchmarks and the results on Cityscapes (70.8%), Camvid (73.2%) and KITTI (51.8%) test sets clearly demonstrate a competitive performance under the real-time category. Among the real-time scene parsing models under 1.5 million parameters, the proposed SDBNet produces the state-of-the-art (SOTA) results on all three datasets.