Lightweight Fire and Smoke Detection with YOLO11: A Two-Benchmark, Multi-Seed Study of Wise-IoU and GhostConv

Vision-based fire and smoke detection must be both accurate and lightweight for edge cameras and unmanned aerial vehicles (UAVs). Most lightweight fire detectors, however, are validated on a single dataset from a single run, which leaves the accuracy–efficiency trade-off and its external validity only partially examined. Rather than a new state of the art, we take an evaluation-centered stance and study two lightweight operating points of YOLO11n: an accuracy-first variant (LFS-YOLO11-A) that adopts the Wise-IoU (WIoU) loss at no extra parameters, and a lightweight variant (LFS-YOLO11-B) that further adds GhostConv. Both are evaluated on the public D-Fire dataset, retrained on a second dataset (DFS) over eight seeds, and profiled across GPU/CPU under PyTorch and ONNX Runtime. LFS-YOLO11-A matches the baseline on D-Fire (mAP@0.5 0.760 versus 0.758), while LFS-YOLO11-B reduces parameters by 12.4% and computation by 11%, both exceeding 160 FPS end-to-end (batch = 1) on a desktop-class GPU. On DFS, WIoU yields a small, exploratory +0.6-point improvement (nominal paired p = 0.037; seed-sensitive, with a confidence interval lower bound near zero), whereas an apparent +6.1-point gain from an early, uncontrolled run proved to be train/test contamination introduced before the data pipeline was frozen, not a genuine effect. Frozen-pipeline, multi-seed, two-benchmark evaluation is therefore necessary to separate genuine effects from the artifacts of uncontrolled single runs in lightweight fire and smoke detection.