Discriminative Cross-Modal Attention Approach for RGB-D Semantic Segmentation

Scene understanding through semantic segmentation is a vital component for autonomous vehicles. Given the importance of safety in autonomous driving, existing methods are constantly striving to improve accuracy and reduce error. RGB-based semantic segmentation models typically underperform due to information loss in challenging situations such as lighting variations and limitations in distinguishing occluded objects of similar appearance. Therefore, recent studies have developed RGB-D semantic segmentation methods by employing attention-based fusion modules. Existing fusion modules typically combine cross-modal features by focusing on each modality independently, which limits their ability to capture the complementary nature of modalities. To address this issue, we propose a simple yet effective module called the Discriminative Cross-modal Attention Fusion (DCMAF) module. Specifically, the proposed module performs cross-modal discrimination using element-wise subtraction in an attention-based approach. By integrating the DCMAF module with efficient channel- and spatial-wise attention modules, we introduce the Discriminative Cross-modal Network (DCMNet), a scale- and appearance-invariant model. Extensive experiments demonstrate significant improvements, particularly in predicting small and fine objects, achieving an mIoU of 77.39% on the CamVid dataset, outperforming state-of-the-art RGB-based methods, and a remarkable mIoU of 82.8% on the Cityscapes dataset. As the CamVid dataset lacks depth information, we employ the DPT monocular depth estimation model to generate depth images.