TIR-FOD: A Thermal-Infrared Benchmark Dataset for Foreign Object Debris Detection on Airfields

What's new in version 1.3 Version 1.3 adds TIR-FOD_v1.3_WEIGHTS.zip: the fine-tuned best-validation weights of all 29 detector runs reported in the article accepted for publication in IEEE Access, with a manifest of run, seed, SHA-256 and recorded test mAP@[.50:.95]. Nine weights match the SHA-256 stored in their run records. The other 20 were re-evaluated on the test split and reproduce their recorded mAP@[.50:.95] to within 0.0004. The dataset files are unchanged from version 1.2. The weights carry the licence stated in LICENCE_NOTICE.txt inside the archive. Dataset TIR-FOD is a 23-class long-wave-infrared (LWIR) dataset and detection benchmark for foreign object debris (FOD) on airfield surfaces. It contains 3,499 single-shot thermal source frames acquired over real runway surfaces at object-to-camera distances of 0.5–10 m under low ambient illumination. These frames contain 5,593 per-instance bounding-box annotations; 281 frames are deliberately object-free. Offline augmentation expands the archival pool to 29,243 images, comprising the source frames and 25,744 stored derivatives. Derivatives are explicitly distinguished from independently acquired images. Images are 640 × 512-pixel, 8-bit, three-channel false-colour JPEGs from a single uncooled vanadium-oxide microbolometer channel operating at 8–14 μm. Their colours encode the camera's thermal display palette, not separate visible-spectrum measurements or calibrated temperatures. The dataset supports appearance-based detection rather than thermometry. Contents The dataset files, unchanged since version 1.2, provide: the complete image archive, original and benchmark-processed annotations, and taxonomy metadata; frozen experimental partitions; structural and palette-dependent annotation audits, cross-split similarity analysis, acquisition-block analysis, and bootstrap confidence intervals; settings, training histories and evaluation results for 29 completed runs covering YOLOv8, YOLO11 and YOLO12 configurations, derivative-handling experiments, and acquisition-block comparisons; portable tools for integrity verification, dataset preparation, result summarization and reproduction of the recorded training configurations. Fine-tuned weights for all 29 runs are included from version 1.3. Three considerations for correct use Preserve the supplied benchmark partitions. The primary clean benchmark contains 18,221 training images, 529 validation originals and 520 test originals. Stored training derivatives originate only from training sources. Four-image mosaics remain in the archive but are excluded from the clean benchmark. Randomly splitting the archival pool can introduce derivative leakage. In a deliberately contaminated, size-matched YOLOv8s experiment, held-out-source derivatives increased mAP@[.50:.95] by 8.52 ± 0.19 percentage points across three seed pairs; this is not a universal estimate for arbitrary splits. Distinguish source separation from acquisition-block separation. The release identifies 117 blocks using consecutive acquisition indices with gaps of at most two. A complementary 12-class experiment separates these blocks across partitions. This evaluates sensitivity to acquisition grouping, rather than establishing cross-session, cross-airport or cross-sensor generalization. Interpret automated annotation checks appropriately. Of the 5,593 source boxes, 5,055 (90.4%) receive localization support from at least one of two automated methods, including 2,824 (50.5%) supported by both. These class-agnostic checks support annotation review; they do not establish semantic correctness or measure human inter-annotator agreement. Licence The dataset is licensed under CC BY-NC 4.0. The weights carry the licence stated in LICENCE_NOTICE.txt. Commercial licensing enquiries concerning the dataset should be directed to the rights holder. Intended use TIR-FOD is intended for research. It has not been validated for operational safety decisions and must not be used for autonomous runway clearance or as a substitute for established airport inspection procedures.