aihpi/bottle-cap-date-stamps

Preussenquelle Bottle-Cap Date Stamps

Photographs of reusable mineral-water bottle caps, each labelled with the date stamp printed on the cap. The dataset was collected to study the reuse lifecycle of returnable glass bottles: The date plus two years (shelf life date) is printed on every cap at filling time, so reading it off returned empties allows the round-trip (fill → return) duration to be measured.

The setup for taking the photographs involved diffusive light to minimize reflections on the aluminium. Dates printed on a inkjet-printed curved, reflective, ridged aluminium surface are challenging for classical OCR. This makes the dataset a useful benchmark for OCR / vision-language models on real-world text.

Dataset structure

A single flat folder: all images at the root, with one metadata file.

.
├── README.md          # this card
├── metadata.csv       # one row per image
├── 260422_001.jpg
├── 260510_001.jpg
└── …                  # 438 images total

File names follow <YYMMDD>_<NNN>.jpg, where the prefix is the capture (return-reading) date and NNN is a per-batch running index.

Data fields

Field	Type	Description
image	image	The cap photograph (loaded automatically by the imagefolder builder).
file_name	string	Image filename (used to join labels to images).
label	string	The date printed on the cap(s), format DD.MM.YY. The physical stamp also includes a time (DD.MM.YY/HH:MM, e.g. 15.04.28/13:40); the time is intentionally excluded. For images containing multiple caps, the dates are joined with ; in left-to-right order (e.g. 03.02.28;16.01.28;31.07.27).
num_caps	int	Number of caps (and therefore dates) visible in the image.

Batches

The data was captured in four sessions. Early sessions photographed one cap per image; later sessions photographed three caps per image to speed up collection.

Batch (prefix)	Capture date	Images	Caps per image	Total date labels
260422	2026-04-22	115	1	115
260510	2026-05-10	116	mostly 3	347
260604	2026-06-04	102	mostly 3	305
260611	2026-06-11	105	3	315
Total	—	438	—	1082

There is a single train split.

Usage

from datasets import load_dataset

ds = load_dataset("<user-or-org>/<dataset-name>", split="train")
example = ds[0]
example["image"]      # PIL.Image
example["label"]      # e.g. "03.02.28;16.01.28;31.07.27"
example["num_caps"]   # e.g. 3

# split a multi-cap label into individual dates:
dates = example["label"].split(";")

Labelling notes and limitations

• One row per image, not per cap. For multi-cap photos the label lists every date in the image (left-to-right).
• Left-to-right ordering of the joined dates reflects the physical order of the caps in the photo.
• Label provenance. Each date was first labeled by a vision-language model. Flagged dates were then checked against the image and corrected by a human reviewer in these cases. Human errors are possible.
• Date format is DD.MM.YY (two-digit year, 20xx).
• Time is excluded by design. Each cap is stamped with both a date and a time (DD.MM.YY/HH:MM). Only the date is transcribed into label. The time portion is visible in the images if you need it.
• No personal data. The images show only bottle caps.

Source and collection

Collected within a pilot project between the AI Service Centre Berlin-Brandenburg (AISC) (Hasso-Plattner-Institut) and Rheinsberger Preussenquelle GmbH. Images were captured with consumer camera equipment under controlled lighting.

Contact

• David Goll
• Leonie Weber

kisz[at]hpi.de

License

Released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share and adapt the data for any purpose, including commercially, provided you give appropriate credit, link to the license, and indicate if changes were made.

See https://creativecommons.org/licenses/by/4.0/ for the full license text.

Acknowledgements

Funded by the German Federal Ministry of Research, Technology and Space (BMFTR) under grant code 16IS22092 (»KI-Servicezentrum Berlin-Brandenburg«). Responsibility for the content lies with the authors.