We propose an efficient and generative augmentation approach to solve the inadequacy concern of underwater debris data for visual detection. We use cycleGAN as a data augmentation technique to convert openly available, abundant data of terrestrial plastic to underwater-style images. Prior works just focus on augmenting or enhancing existing data, which moreover adds bias to the dataset. Compared to our technique, which devises variation, transforming additional in-air plastic data to the marine background. We also propose a novel architecture for underwater debris detection using an attention mechanism. Our method helps to focus only on relevant instances of the image, thereby enhancing the detector performance, which is highly obliged while detecting the marine debris using Autonomous Underwater Vehicle (AUV). We perform extensive experiments for marine debris detection using our approach. Quantitative and qualitative results demonstrate the potential of our framework that significantly outperforms the state-of-the-art methods.