Reproduce “Image compositing is all you need for data augmentation” on OpenResearch

About this paper

Ang Jia Ning Shermaine, Michalis Lazarou, Tania Stathaki

This research from Imperial College London and the University of Surrey introduces an Image Compositing method for data augmentation, which achieved a [email protected] of 0.911 for ground-based aircraft detection. The method demonstrably outperformed state-of-the-art generative models like Stable Diffusion XL with ControlNet (0.854 [email protected]) on a custom, limited dataset.

Key insights

• For domain-specific object detection tasks with limited data, a targeted and carefully designed image compositing augmentation strategy can be more effective than advanced general-purpose generative AI models.
• General-purpose generative models, despite their sophistication, can introduce a 'distribution shift' between synthetic and real data, potentially limiting their performance gains in specific niche domains.
• Maintaining high realism and relevance to the target task through controlled compositing, using real foregrounds and backgrounds, is crucial for effective synthetic data generation.

Summary by alphaXiv.