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

v2 — revised annotations (12 August 2026) This version replaces the bounding-box annotations of v1. The 11-class taxonomy and the official train/validation split are unchanged. Results obtained on v1 and on v2 are NOT comparable, and anyone who downloaded v1 should treat its annotations as superseded. What changed: Annotated instances: 3,084 → 13,313 Objects per image: mean 2.06 → 8.88, median 1 → 6 Median normalised box area: 0.025 → 0.004 The first annotation pass was systematically incomplete: 67.7% of images carried exactly one bounding box, which is inconsistent with the cluttered, occluded scenes this dataset is meant to represent. The whole set was therefore re-annotated under a written protocol. The original boxes were created manually in makesense.ai. For v2 those labels were expanded through model-assisted re-annotation: detector outputs were treated as review candidates rather than accepted as ground truth. Candidates were reviewed by a human, incorrect boxes were removed or reclassified, and automated integrity checks were applied to every label file. The validation split was reviewed first and in full, before any training image was touched, so that the evaluation set is independent of the model used to accelerate the rest. 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. Baselines retrained on the revised labels, official validation split, 960 px input, given as Precision / Recall / mAP@0.5 / mAP@0.5:0.95: YOLOv8n — 0.801 / 0.693 / 0.762 / 0.564 YOLOv8s — 0.796 / 0.703 / 0.747 / 0.560 Faster R-CNN MobileNetV3-320 — 0.567 / 0.356 / 0.490 / 0.329 The archive also ships the trained YOLOv8s weights, the training and evaluation scripts, and the annotation protocol. 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 1,500 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.