다양한 극한 환경에서 강인한 의미론적 분할을 위한 소수의 추가 레이어에 관한 연구

Abstract

Semantic segmentation has seen significant progress, achieving high performance under clean weather condition. Despite this, a common limitation among many mod- els is their reduced effectiveness in adverse weather conditions, including rain, dark- ness, snow, and fog. While there have been concerted efforts to address these issues individually, comprehensive studies aiming to tackle all these conditions simultane- ously are relatively scarce. A few endeavors that have ventured into this territory have encountered a notable increase in computational costs, compared to conven- tional methods. Our work introduces a novel approach, effectively addressing the challenges posed by multiple weather conditions, while maintaining a computational efficiency superior to existing methods. This is achieved by identifying layers within the backbone network that are sensitive to changes in weather conditions and desig- nating these as weather adaptive layers. The model selectively combines the outputs of these weather adaptive layers based on the weather condition. The method utilizes a smaller set of parameters yet outperforms conventional approaches. This under- lines the contribution of our approach, demonstrating the feasibility of achieving high performance in semantic segmentation across diverse and challenging weather condi- tions, without the necessity of a substantial increase in computational resources.