Search papers, labs, and topics across Lattice.
This paper addresses the challenge of small object detection in satellite imagery across different geographical regions by using generative AI to create synthetic training data. A Stable Diffusion model is fine-tuned on both source (Selwyn, New Zealand) and target (Utah, USA) regions, leveraging cross- and self-attention mechanisms and CLIPSeg for image segmentation. The approach demonstrates a 20% improvement in detection accuracy on the target dataset compared to a baseline trained solely on source data, highlighting the effectiveness of generative data augmentation for cross-regional generalization.
Synthetic data generated by fine-tuning Stable Diffusion on multi-region satellite imagery boosts small object detection accuracy by 20%, even when real labeled data is scarce.
Detecting small objects like cars in satellite imagery is challenging, especially when labeled data is unavailable for a target region. This study introduces a novel approach using generative AI to address this issue by fine-tuning a Stable Diffusion model on images from one geographical region, the source (Selwyn, New Zealand), and a second region of interest, the target (Utah, USA) environment. Our framework employs state-of-the-art cross and self-attention mechanisms alongside the CLIPSeg image segmentation method to generate high-quality synthetic datasets with minimal supervision. Our empirical results show a significant improvement in detection performance, improving the accuracy by over 20% when testing the target dataset compared to the performance of a baseline model trained only on source data.
CMU ML