Layout-Controlled Synthetic Data Generation for Remote Sensing Object Detection

High-performance remote sensing object detection typically requires vast datasets with annotations, but the collection and labeling of large-scale data is a costly and labor-intensive process. While generative models have shown promise in addressing the need for large-scale datasets, they heavily rely on high-quality labeled data and often overlook crucial aspects, such as the inter-object relationships and overall quality of the synthetic data. To this end, we propose Layout-controlled Synthetic Data Generation (LSDGen), a novel approach for generating high-fidelity remote sensing imagery for object detection. LSDGen consists of three key stages: (1) Relation-Constrained Layout Generation (RLGen), which constructs a category association matrix from benchmark data to generate both main and subcategories, as well as their spatial configurations, without relying on bounding-box annotations; (2) Image Synthesis via Layout-Conditional Diffusion Model, where the generated layout guides a diffusion model to produce high-quality remote sensing images; and (3) Instance-Centered Synthetic Data Filtering (ISDF), which evaluates the semantic consistency of individual instances in the synthesized images by calculating instance-level cosine similarity scores between visual and textual representations, filtering out low-confidence samples based on the average semantic score. Experimental results on the DIOR, DIOR-R, and HRSC datasets demonstrate that our method significantly improves the detection performance of existing remote sensing image object detection models, outperforming state-of-the-art annotation-dependent synthesis approaches.