CACP: Context-Aware Copy-Paste to Enrich Image Content for Data Augmentation

Qiushi Guo, Shaoxiang Wang, Chunpeng Chang, Jason Rambach; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 5186-5195

Data augmentation is a widely used technique in deep learning, encompassing both pixel-level and object-level manipulations of images. Among these techniques, Copy-Paste stands out as a simple yet effective method. However, current Copy-Paste approaches either overlook the contextual relevance between source and target images, leading to inconsistencies in the generated outputs, or heavily depend on manual annotations, which limits their scalability for large-scale automated image generation. To address these limitations, we propose a context-aware approach that integrates Bidirectional Latent Information Propagation (BLIP) for extracting content from source images. By aligning the extracted content with category information, our method ensures coherent integration of target objects through the use of the Segment Anything Model (SAM) and YOLO. This approach eliminates the need for manual annotation, offering an automated and user-friendly solution. Experimental evaluations across various datasets and tasks demonstrate the effectiveness of our method in enhancing data diversity and generating high-quality pseudo-images for a wide range of computer vision applications.