Dual Gated Fusion Network with Feature Calibration for Semantic Segmentation

Recently, deep learning-based feature fusion has attracted extensive interest due to its powerful representation and production capabilities. The goal of feature fusion is to extract more accurate and useful features by processing the relationships between features from a single or several sources, resulting in a more thorough knowledge of the scene. However, Traditional fusion approaches do not retain effective information well and consider the noise in the modalities. Therefore, this paper proposes Dual Gated Fusion Network with Feature Calibration (FC-DGFNet) for semantic segmentation. It mainly consists of two parts: Feature Calibration Module (FCM) and Gated Fusion Module (GFM). FCM is embedded to calibrate RGB and depth features with each other to generate more robust features. And GFM is used for the fusion of depth features and visible features, which improves the representativeness of features and better retains complementary information. Experimental results demonstrate that the proposed method outperforms the traditional methods, and mIoU and PA reach 51.6% and 77.5%, respectively.