MEOBDET: Mendel Everyday Object Detection dataset - 154,854 annotated images of 147 everyday-object categories (version 1.1)

MEOBDET (Mendel Everyday Object Detection) is a bounding-box object-detection dataset of everyday objects: 154,854 images, 739,481 boxes and 147 categories (jewellery, hand tools, electronics, cosmetics, kitchenware, banknotes, board-game pieces, stationery, a few fresh foods and plant leaves), photographed or filmed as isolated objects by 66 consenting contributors (Mendel University in Brno, 2025) and annotated by them in CVAT. COCO-format annotations with training, validation and test partitions (124,416 / 15,318 / 15,120 images); pseudonymous contributor directories data_001 to data_070 keep the contributor grouping visible. Version 1.1 (30 October 2026) corrects two mislabelled collections found by a stratified annotation check (data_009: shifted category ids, 1,695 backpack boxes restored; data_012: 796 pocket-calculator boxes separated from the football-boot class into the new category kalkulacka), corrects 73 boxes of data_038 after a full check of that collection (boxes around passers-by, traffic signs, bags and shoes removed; swapped classes, oversized and missing boxes fixed; images unchanged), and corrects 15,952 further boxes in 7,163 images after a dataset-wide detector check (4,687 duplicate boxes removed, boxes added for missing objects and for 1,476 images that had none, boxes around several objects split, swapped categories corrected, and 1,175 rotated rectangles whose angle was lost at consolidation restored from the exports; 5,104 edits were viewed one by one under a written rubric (Claude models, the author checked samples), 432 were decided by a judge trained on those decisions (held-out precision 96 %), the systematic duplicates and rotations were checked on random samples, and the 5,588 boxes of unannotated video frames come from a detector trained on the same videos; every edit and its source is listed in the audit archive), regenerates the face anonymisation with an annotation-aware rule (the version 1.0 pass had pixelated 8.6 % of the annotated objects), moves every image into the split folder of its partition, removes unvalidated masks from the main files, and adds a top-level README, English glosses and thematic groups for all categories (categories.csv), a per-image manifest with hashes and duplicate and anonymisation flags (images_manifest.csv.gz), cleaned validation and test lists, the complete audit outputs (MEOBDET_audit.tar.gz), a 300-image preview with drawn boxes (MEOBDET_preview.zip) and conversion and evaluation tools (MEOBDET_tools.tar.gz). Details in README.md and CHANGELOG.md. Download: extract MEOBDET_annotations.tar.gz and the 69 MEOBDET_data_NNN.tar archives (data_038 comes in four parts) into one folder; image paths are the file_name values of the COCO files; verify with SHA256SUMS.txt. Known limitations (read README section 8): each category comes from one contributor and one environment, so the supplied partitions measure within-collection recognition only; frames of the same video occur in all partitions (use the cleaned lists); annotation conventions differ between contributors. Licence CC BY-NC 4.0, aligned with the scope of the written consent (22 October 2025). Version 1.0 (3 July 2026) remains available under its own DOI.