ADAAUG: An Adaptive Data Augmentation Method for Change Detection

Current data augmentation methods for change detection utilize fixed strategies to produce image samples without considering the specificity of the change image pairs. In this paper, we propose ADAAUG, a new reinforcement learning framework that adaptively augments the change image pairs. An actor selects the best augmentation operation from the operation set according to the image pair. The augmented image pairs make the change detector easier to learn the optimal parameters and improve the final detection performance. Besides, we propose a mask guided mixing customed for change detection data augmentation, which mixes the change regions of the current training sample based on the prediction results and labels to generate high-quality samples with more positive samples. The extensive experiments on the standard benchmarks show that ADAAUG achieves favorable performance compared to SOTA data augmentation methods.