DBiSeNet: Dual bilateral segmentation network for real-time semantic segmentation

Bilateral networks have shown effectiveness and efficiency for real-time semantic segmentation. However, the single bilateral architecture exhibits limitations in capturing multi-scale feature representations and addressing misalignment issues during spatial and contextual feature fusion, thereby constraining segmentation accuracy. To address these challenges, we propose a novel dual bilateral segmentation network (DBiSeNet) that incorporates an additional bilateral branch into the original architecture. The additional (high-scale) bilateral operating at high resolution to preserve fine-grained details and responsible for thin object prediction, while the original (low-scale) bilateral maintains an enlarged receptive field to capture global context for large object segmentation. Furthermore, we introduce an aligned and refined feature fusion module to mitigate feature misalignment within each bilateral branch. To optimize the final prediction, we design a dual prediction fusion module that utilizes the low-scale segmentation results as a baseline and adaptively incorporates complementary information from high-scale predictions. Extensive experiments on the Cityscapes and CamVid datasets validate the effectiveness of DBiSeNet in achieving an optimal balance between accuracy and inference speed. In particular, on a single RTX3090 GPU, DBiSeNet2 yields 75.6% mIoU at 225.9 FPS on Cityscapes test set and 75.7% mIoU at 203.4 FPS on CamVid test set.