MEOBDET: A 154,854-Image, 146-Category Object Detection Dataset Built from Consented, Anonymised Student Contributions

MEOBDET (Mendel Everyday Object Detection) is a 154,854-image, 733,802-bounding-box object detection dataset spanning 146 everyday-object categories -- jewellery, workshop tools, electronics, cosmetics, kitchenware, banknotes, board games, and a smaller set of food items and plant leaves photographed as isolated objects, not agricultural or field imagery. Images and COCO-format annotations were collected by 66 anonymised participants (of 70 who consented; data_018/026/029/033 contributed no annotations) of an artificial-intelligence course at Mendel University in Brno, under written per-group consent (dated 22 October 2025) explicitly authorising open-source publication for scientific and research purposes. Provided with standard train/val/test COCO splits (124,416 / 15,318 / 15,120 images) and a parallel YOLO dataset.yaml. RF-DETR fine-tuned on the full training split reaches 94.4-95.3% mAP@50 across three model sizes; full technical validation, construction methodology, and known data quirks are documented in the accompanying Data in Brief article and README. Anonymisation verified: numeric file names, re-mapped COCO file_name fields, and a full byte-level scan confirming no embedded EXIF/GPS/metadata across all 154,854 images. Licensed CC BY-NC 4.0. Contributors listed on this record participated in data collection and basic annotation of their individual subsets. The final dataset assembly, cleaning, harmonisation, annotation correction, quality control, anonymisation, documentation, release preparation and model training/validation were performed by Miroslav Jaroš.