ImDUSTRY5-1.5K: A Public Benchmark Dataset for Industrial Object Detection in Real Production Environments

v3 — anonymised release (September 2026) Use this version. v1 and v2 contain the images before anonymization, and v1 also carries a superseded, incomplete annotation pass. What changed from v2: – Faces that appeared incidentally were irreversibly blurred in 33 images: a printed photograph of a young child pinned on a workstation pegboard, visible in 29 frames of the eye-tracker scene video, and a person seen through a window in the background of four photographs. No blurred region overlaps an annotation box, so no label changed.– The 33 PNG files were removed together with their 445 annotation boxes. They were screen captures of processed eye-tracker scene video, several with detector prediction boxes, hand-landmark overlays or video-player controls rendered into the pixels.– Images: 1,500 → 1,467 (1,240 training, 227 validation). Annotation boxes: 13,313 → 12,868.– The three baselines were retrained on v3.– Added audit_labels.py, which repeats the label integrity checks, and docs/CHANGES_v3.md with the lists of removed and anonymised files. Baselines on the official validation split, as Precision / Recall / mAP@0.5 / mAP@0.5:0.95. YOLO models at 960 px input, with precision and recall at the confidence threshold that maximises mean F1; Faster R-CNN at 320 px input, with precision and recall at a score threshold of 0.25. YOLOv8n — 0.780 / 0.732 / 0.783 / 0.563YOLOv8s — 0.784 / 0.748 / 0.763 / 0.562Faster R-CNN MobileNetV3-320 — 0.631 / 0.323 / 0.499 / 0.321 The images come from two sources: 976 smartphone photographs and 491 frames from the scene camera of a Pupil Labs Neon head-mounted eye tracker. Only scene-camera frames are released; no eye-camera images or gaze data. docs/ANNOTATION_PROTOCOL.md inside the archive documents the annotation rules, including class boundaries that are not obvious from the class names — in particular, "Wrench" denotes the powered nutrunner used at this workstation, not a hand tool for fasteners. The archive ships the images, the YOLO-format labels, the official split, the annotation protocol, the trained YOLOv8s weights, and the training, evaluation and label-audit scripts. Code and issue tracker: https://github.com/Kanan02/YoloTraining ImDUSTRY5-1.5K is a public benchmark dataset for industrial object detection collected in a real production environment collected under the project HABIT. The dataset contains 1,467 annotated images across 11 object classes and is designed for reproducible evaluation under realistic industrial conditions, including clutter, occlusion, scale variation, and non-uniform illumination. The record includes YOLO-format annotations, an official train/validation split, metadata files, and documentation.