In fields such as medical diagnosis and intelligent driving, there is a growing demand for accurate instance segmentation. This paper mainly reviews data augmentation techniques and their impact on the performance of instance segmentation models. The purpose of this study is to evaluate the effectiveness of different augmentation strategies in alleviating data insufficiency and improving model generalization. This paper uses a document analysis method to summarize the results of image transformation-based (geometry, color, filters, random erasure, image blending) and depth network-based (GAN, neural style transfer) enhancement methods. The results show that geometric and color transformations enhance the model's adaptability to lighting variation, random erasing and image mixing enhance its robustness to occlusion and complex backgrounds, and synthetic samples generated by GAN and neural style transfer further optimize segmentation accuracy. The study found that a reasonable combination of multiple augmentation strategies can significantly improve instance segmentation performance, providing practical references for related applications and algorithm improvements.
Yana Yang (Tue,) studied this question.