moneebhussain/RegionalPack-VLM-Dataset

RegionalPack-VLM: Retail Packaging Recognition Evaluation Dataset

Researcher: Moneeb Hussain
Version: 0.1 | October 2026
Location of collection: Lahore, Pakistan
Dataset size: 60 photographs of 20 retail products

RegionalPack-VLM is a small, manually annotated image evaluation dataset I collected in Lahore, Pakistan to investigate product-line identification, packaging-text transcription, and category recognition in vision-language models.

It contains 60 smartphone photographs of 20 branded retail products, each captured straight-on, tilted, and partially occluded. I manually verified the front-of-pack text annotations for every photograph.

Nine international/Pakistan-market category pairs, representing 18 products, form the primary comparison. One additional apple-drink pair is exploratory. The collection is intended for transparent evaluation and inspection of model behavior, not as a representative benchmark of Pakistani retail products.

Dataset Composition

Twenty products, three photographs each. The primary categories, one internationally distributed product and one Pakistan-market product each, are cola, biscuit, tea, chocolate, chips, detergent, toothpaste, noodles, and soap. The apple-drink pair, apple_fruita and apple_fruitien, is exploratory.

Each row of metadata.csv points to one JPEG in images/. The filename is <product_id>_<angle>.jpg. The file_name value is the path from this dataset root, for example images/cola_intl_straight.jpg.

The Hugging Face image loader names this single collection train. That name is the library default for a folder that does not define splits. It is not a training partition. This export does not define validation or test splits.

Collection Methodology

I photographed the products in Lahore with a phone. The three conditions straight, tilt, and occluded name how I held the phone. They are not a controlled viewpoint or robustness experiment. I am the only photographer and the only annotator.

Annotation Schema

brand_true is the specific product line. It is not the parent company. YOU is the product line on the CandyLand bar. Rite is the product line on the Bisconni biscuit. CandyLand and Bisconni are not stored as brand_true.

comparison_role is copied from the product manifest, where I recorded it. It is primary for 18 products and exploratory for the two apple drinks. The photograph annotation file does not contain that column.

Field	Meaning
file_name	Path of the JPEG in this dataset
product_id	Product identifier
angle	straight, tilt, or occluded
comparison_role	primary or exploratory, from the product manifest
origin	international, local, or exploratory
brand_true	Specific product line
text_true_latin	Verified front-of-pack Latin text
text_true_urdu	Verified Urdu text. Empty when none is in scope
review_status	verified on all 60 rows
category	Product category, such as cola or biscuit
variant	Short description of the photographed variant
brand_group	Collection group: international, pakistan_origin, or exploratory
script_present	Script recorded for the scored reference
occlusion_note	Note on what the frame hides, when I wrote one
brand_visible	Whether the product-line name is visible in that frame
ambiguous_text	Text I saw but did not put in the scored reference
notes	Other annotation notes
image_path	Original path in the research repository, data/raw/<product_id>/<angle>.jpg

Ground-Truth Verification

I verified the front-of-pack text on all 60 photographs before any model was run. The scored reference keeps the visible brand wordmark, the product or flavor name, and a prominent tagline. Prices, nutrition tables, barcodes, and regulatory fine print are outside the reference.

Primary and Exploratory Comparison

There is no training split. All 60 images are an evaluation set.

Comparison	Products	Photographs	How I use it
Primary	18	54	The paired comparison
Exploratory	2 apple drinks	6	Kept in the collection and excluded from the primary comparison

Intended Uses

A researcher can inspect the photographs and these annotations to study packaging transcription, category recognition, and product-line identification on this shelf.

The set is small and selected. It is not representative of Pakistani retail products. It is not a basis for a causal claim about geographic bias. There is no training split. A result computed after training on these same images would not be a held-out result.

Limitations

One annotator. Eighteen primary products. Pairs share a category and are not matched variants. Urdu references exist for 9 photographs. Tilt and occlusion are informal phone conditions.

Licensing and Intellectual Property

I took the photographs. Brand names and pack designs are trademarks of their owners. This card does not grant rights to those marks. The license: other field means this dataset is not released under a standard open license. The photographs, annotations, and accompanying text are not placed under an open license. Model weights are not included.

Related Resources

Resource	Where
Research repository	RegionalPack-VLM on GitHub
Photographs, responses, and technical report	RegionalPack-VLM Bucket
Evaluated model	Qwen3-VL-2B-Instruct

The technical report is reports/technical_report.md in the GitHub repository and in the Bucket. This dataset repository contains the photographs and the verified metadata only.