CroPe: Cross-Modal Semantic Compensation Adaptation for All Adverse Scene Understanding

Scene understanding in adverse conditions, such as fog, snow, and night, is challenging due to the visual appearance degeneration. In this context, we propose a Cross-modal Semantic Compensation Adaptation method (CroPe) for scene understanding. Distinct from the existing methods, which only use the visual information to learn the domain-invariant features, CroPe establishes a visual-textual paradigm which provides textual semantic compensation for visual features, enabling the model to learn more consistent representations. We propose the Complementary Perceptual Text Generation (CPTG) module which generates a set of multi-level complementary-perceptive text embeddings incorporating both generalization and domain awareness. To achieve cross-modal semantic compensation, the Reverse Chain Text-Visual Fusion (RCTVF) module is developed. By the unified attention and reverse decoding chain, compensation information is successively fused to the visual features from the deep (semantic dense) to shallow (semantic sparse) features, maximizing compensation gain. CroPe yields competitive results under all adverse conditions and significantly improves the state-of-the-art performance by 6.5 mIoU for ACDC-Night dataset and 1.2 mIoU for ACDC-All dataset, respectively.