Class-Specific Noise Injection for Improved Road Segmentation
摘要
In this paper, we introduce a novel class-specific noise method designed for efficient data augmentation in the realm of road segmentation. This approach is rooted in the observation that in practical image segmentation, edges area of specific class often holds higher level of importance than interiors. Distinct from traditional data augmentation techniques, our method tailors the generation of noise based on the specific class. Through experimental validation, we demonstrate that our proposed approach can significantly bolster the mean intersection over union (miou) performance of models on test datasets. Our technique holds potential for a broad spectrum of image segmentation tasks, including but not limited to medical imaging and road segmentation.