SimD3: Simulated Drone Detection Dataset

Dataset Organization The SimD3 dataset contains a total of 1,78,639 synthetic RGB images, rendered at a fixed resolution of 1920 × 1080 pixels. The dataset is organized into three primary subsets, each released with predefined training, validation, and test splits: Non-VFX subset:112,899 images, where both drones and birds are explicitly annotated.This subset supports supervised learning and evaluation for drone–bird discrimination. VFX subset:46,086 images, where drones are annotated and bird-like motion is introduced using Unreal Engine Niagara visual effects.Birds in this subset are intentionally treated as dynamic background clutter and are not annotated. Weather subset:19,654 images, rendered under adverse weather conditions such as fog and snow, with full annotations for drones and birds. Across the entire dataset, images include frames containing only drones, only birds, and scenes where drones and birds appear simultaneously. Each archive follows the directory structure: images labels All annotations are provided in YOLOv5 format. Classes The dataset includes the following classes: 0: drone 1: bird Note: In the VFX subset, birds appear visually but are not annotated and therefore do not appear in label files. Image Properties Resolution: 1920 × 1080 pixels Minimum object size: ~5 pixels Maximum object size: up to 20% of image area Payload Modeling Drone images with attached payloads can be identified using filename prefixes. The naming convention encodes drone type and payload: H, Q, O denote hexacopter, quadcopter, and octocopter platforms BAG, BOX, GUN indicate payload types Intended Use SimD3 is intended for academic research in drone detection, small-object detection, robustness analysis, domain generalization, and sim-to-real transfer. The dataset is suitable for training and evaluating real-time object detection models under diverse environmental and operational conditions.