Synthetic Aerial Dataset for UAV Detection via Text-to-Image Diffusion Models

In this work, we present an approach to generate a synthetic aerial dataset for efficient Unmanned Aerial Vehicle (UAV) detection. We propose controlling the output of a text-to-image diffusion model by applying additional input conditions. Specifically, we train a diffusion model that enables conditional inputs, i.e., binary masks that specify all tractable parameters, including quantity, scale, pose, color, background, etc. Diverse photorealistic images with corresponding ground truth bounding boxes are generated automatically in an end-to-end manner. Without any interference, the dataset can be scaled to a large magnitude to facilitate the training process of UAV detection. Experimental results of YOLOv7 trained on the synthetic dataset demonstrate an extensive precision increment on unseen datasets of real images.