Evidence-Based Fusion for Low-Quality RGB-T Semantic Segmentation

In recent years, significant progress has been made in RGB-thermal (RGB-T) semantic segmentation. However, most methods assume optimal imaging conditions, overlooking real-world challenges, such as sensor noises, blur, extreme weather, or transmission errors. These factors can lead to inconsistent information across RGB and thermal modalities, degrading segmentation performance. To address this, we propose an evidence-based fusion network (EBFNet) to enhance segmentation accuracy for low-quality RGB-T inputs. Concretely, our approach tackles inconsistent modal information by introducing an evidence-based consistency loss (ECL), which quantifies and mitigates conflicts between modalities using a Dirichlet distribution to model decision reliability. Building on this, we design a discrepancy priority fusion module (DPFM). It fuses RGB and thermal features by prioritizing reliable ones using discrepancy maps. This enhances feature integration, boosting robustness for low-quality inputs. For evaluation, we construct low-quality versions of the MFNet and PST900 datasets by simulating real-world degradations. Extensive experiments on these datasets demonstrate that EBFNet outperforms state-of-the-art (SOTA) methods in low-quality RGB-T semantic segmentation.