Drone-Helper: cross-dataset YOLOv8 fire and smoke detection — trained models, evaluation outputs and audit artefacts

Drone-Helper: cross-dataset YOLOv8 fire and smoke detection — trained models, evaluation outputs and audit artefacts Companion deposit for the manuscript: Ruangsang, W.; Pramkeaw, P. Cross-Dataset Evaluation of YOLOv8 for Unmanned Aerial Vehicle Fire and Smoke Detection: Benchmark Contamination, Zero-Shot Transfer, and Onboard Deployment on a Low-Cost Airframe. Submitted to Drones (MDPI), manuscript drones-4441317. Contents drone-helper-weights.zip — the twelve trained checkpoints (best.pt), one per run, named <corpus>__<variant>__seed<k>: YOLOv8n/s/m x seeds 0-2 on D-Fire, YOLOv8n/s/m x seed 0 on FASDD_UAV. Each run folder also carries its complete Ultralytics configuration (args.yaml) and per-epoch training log (results.csv). drone-helper-results.zip — the raw evaluation tables behind every number in the paper: results_indomain.csv (validation and official-test-split metrics, Table 6), results_crossdataset.csv (the 2x2 zero-shot matrix, Table 7), results_dedup.csv (full / clean / dup evaluations, Table 9), and environment.json (Python, PyTorch 2.2.2+cu121, CUDA 12.1, cuDNN 8902, Ultralytics 8.4.70, GPU, and the complete pip freeze). drone-helper-dedup.zip — the near-duplicate contamination audit and its correction: per-image 64-bit perceptual hashes for every train and test image of both corpora (phash_cache/), the exact clean/dup test image lists and data yamls (subsets/), the evaluation script (eval_dedup.py), and the audit reports (Tables 8 and 9 of the paper). drone-helper-datasets-scripts.zip — dataset provenance and rebuild tooling: the unmodified official split-definition files, split provenance notes including the FASDD_UAV class-index remap (0=smoke, 1=fire), dataset build scripts that verify counts and abort on mismatch, the audit script, and the training and evaluation driver (train_plan_rtx3060.py). Source imagery (not redistributed here) D-Fire (21,527 images, CC0 1.0): github.com/gaia-solutions-on-demand/DFireDataset — archive used: D-Fire.zip, 3,036,222,313 bytes. FASDD_UAV (25,097 images, CC BY-SA 4.0): Science Data Bank, https://doi.org/10.57760/sciencedb.j00104.00103 — archive used: FASDD_UAV.zip, md5 c5ea9651ca672fc128c6f17d0797c2f2. Licence CC BY-SA 4.0. The FASDD_UAV-trained checkpoints derive from data released under CC BY-SA 4.0; the share-alike term is therefore applied to the whole record. Attribute FASDD to Wang et al. (Geo-Spatial Information Science 28(2):511-526, 2025) and D-Fire to de Venancio et al. (Neural Computing and Applications 34:15349-15368, 2022). Reproducing Download both corpora from their original sources above. Rebuild the exact splits: python scripts/build_dfire.py, python scripts/build_fasdd.py (both verify image counts and label indices, and abort on mismatch). Retrain or evaluate: python scripts/train_plan_rtx3060.py --check then --profile lean; the cross-dataset matrix alone: --matrix-only. Contamination-corrected evaluation: python dedup_analysis/eval_dedup.py.