OmniSearch: Synthetic Multi-Modal Survivor-Detection Datasets for Wildfire Search & Rescue

Physics-grounded synthetic datasets and trained YOLOv8 detectors for survivor detection in wildfire search-and-rescue, produced for the UC Berkeley MIDS OmniSearch capstone. Includes aerial (drone) datasets with altitude-aware sizing and oblique (side-angle) camera views, ground-robot (UGV) datasets for front and mast cameras, and a simulated thermal-infrared (TIR) dataset. Real person cutouts (SARD) and hard-negative vehicle crops (VisDrone) are composited over NAIP aerial imagery with color harmonization, range-aware blur, and wildfire effects. Bundled with the generation scripts, a datasheet (EDA), and trained model weights for full reproducibility. Drone survivor (SARD+NAIP composite): 2,000 train + 400 val (25% oblique) Drone survivor (NAIP-only background): 500 train + 90 val UGV front + mast cameras: 3,000 train + 600 val Thermal TIR: 1,000 train + 200 val Source assets: 54 SARD cutouts, 500 VisDrone decoys, 81 NAIP tiles Labels are in standard YOLO format (class 0 = person) with per-image metadata JSONs.