EVAFusion: Environment-Aware Fusion With Conditional Modulation for Multimodal Segmentation
Zhihao Zheng, Limin Xiao, Ming Zhao, Yunzhou Li · IEEE Access · 2026
Multimodal perception improves robustness by integrating complementary sensor information, yet most existing approaches rely on static fusion strategies and assume favorable environmental conditions, leading to performance degradation in adverse scenarios. We propose an Environment-aware Adaptive Multimodal Perception framework that explicitly models environmental conditions and dynamically adjusts multimodal fusion. An environment condition recognition module is used to guide fusion via condition-aware modulation. Our fusion design combines window-based cross-attention for efficient local interactions, AdaLN-based conditional feature modulation for continuous modality reweighting, and Conditional Positional Encoding (CPE) for input-dependent spatial alignment. This enables condition-aware perception across diverse environments while maintaining a compact shared-backbone architecture. Extensive experiments on MUSES and DELIVER, together with condition-wise, degradation, efficiency, and ablation analyses, show that the proposed framework improves segmentation performance in several adverse conditions while also revealing clear limitations under strong image noise and trade-offs in several auxiliary design choices.