shivank21/mmconflict-restaurant-menus-500

MMConflict Indian and international restaurant menus

This dataset contains 1,000 restaurant menu images. It has 500 Indian menu cards and 500 international historical menus. The menu_region column contains indian or international.

menu_region describes the source collection. It is not a model prediction about a restaurant, cuisine, language, or country.

Sources and licenses

Each international row contains its Commons page, NYPL menu ID, menu page ID, image ID, structured menu item ID, location when available, and file-level license metadata. The released international images have a maximum side of 1,600 pixels. Their bounding boxes were scaled to the released image size.

Prices, boxes, and questions

The Indian rows use Tesseract OCR candidate prices and item names. The international rows join NYPL's structured dish and price records to the matching menu page, then use Tesseract to locate that known price in the image.

Questions use What price is printed for {item_name}?. editable_bbox uses [x, y, width, height] in final JPEG pixels.

Distinctness

All 1,000 rows have different image files and pixel hashes. A 64 bit difference hash computed from the released JPEGs has no pair within Hamming distance 4.

The Indian subset was selected with ResNet 18 embeddings, greedy farthest point selection, nearby source ID filtering, and a cosine limit of 0.93. The international subset was selected from 683 strict candidates with farthest point embedding selection and a cosine limit of 0.98. It contains one page from each of 500 different NYPL physical menu IDs. ResNet 18 nearest neighbor values were recomputed across all 1,000 released images and are stored in each row. Summary values are in distinctness_report.json.

Limitations

• —The price, item name, question, and bounding box annotations were generated automatically and were not checked by a person. Verify them before using the rows as gold evaluation data.
• —The images are distinct, but some Indian rows may be different pages, crops, or photographs of the same restaurant menu. The international rows use one page per NYPL physical menu ID.

Main fields

• —image is the menu image.
• —menu_region is indian or international.
• —source_id, source_url, source_license, license_url, and attribution describe the source.
• —item_name, visible_value, normalized_value, and currency describe the candidate target.
• —editable_bbox, bbox_format, bbox_source, ocr_engine, and ocr_confidence describe localization.
• —nearest_image_id, nearest_selected_cosine_similarity, perceptual_hash, and sha256 support distinctness checks.

Suitable conflicts

Rows can support old and new price conflicts, restaurant and delivery price differences, weekday or location variants, currency changes, size dependent prices, availability changes, and OCR or editing errors.