Boosting Generalization of Semantic Segmentation With Unseen Style Seeking-Based Meta-Learning.

Summary: Imagine teaching a computer to recognize cars, trees, and roads using only pictures taken on sunny days. If you then show it a picture taken in the rain or snow, it usually gets confused! This paper introduces a clever new tool called USSML. It takes that one sunny picture and automatically imagines what it would look like in many different styles—like rainy, snowy, or painted. By training on these "imagined" pictures, the computer learns to recognize objects no matter what the picture looks like. It works great, doesn't need much extra computer power, and beat older methods in tests using five different sets of pictures.