Multimedia-SMU/multilingual-image-recipe-retrieval

Description

This is the first real-world multicultural cuisine dataset designed for the image–recipe retrieval task. It addresses limitations of existing datasets by providing multilingual recipes spanning five Southeast Asian cuisines (Indonesia, Malaysia, Thailand, Vietnam and India).

Motivation

Existing image–recipe retrieval datasets focus primarily on monolingual settings, whereas in real-world applications, recipes originate from diverse regions worldwide and are described in multiple languages and cultural contexts. Our dataset addresses this limitation by collecting multilingual recipes from five Southeast Asian regions, enabling more realistic and comprehensive image–recipe retrieval research.

Dataset Curation and Statistics

The image–recipe pairs are crawled from Cookpad, using dish titles compiled from Wikipedia, including lists of Indonesian, Malaysian, Thai, Vietnamese, and Indian.

Given a dish title, a ranked list of image–recipe pairs (ranging from 1 to 6,469 pairs) is retrieved. For example, recipes such as "nasi lemak ipin upin", "nasi lemak hijau pandan", and "sambal nasi lemak" are obtained using "nasi lemak" as the search keyword. In total, we curate a dataset containing 104,842 image–recipe pairs, collected using 776 dish titles across five different cultures.

To ensure data quality, we perform deduplication by removing samples with duplicate recipe titles in the test set. Specifically, for each group of duplicate recipes, we randomly retain one sample and discard the rest. The statistics of the training, validation, and test splits are summarized in the following table.

Cuisine	Train	Val	Test
Indonesia	18,001	3,177	3,588
Malaysia	13,099	2,312	3,437
Thailand	16,833	2,971	3,977
Vietnam	15,045	2,656	3,145
India	10,618	1,874	4,109
Total	73,596	12,990	18,256

Dataset Statistics

The following table shows the distribution of the dataset across different cultures, which is notably imbalanced. Indonesia has the highest number of query keywords and consequently the most crawled image-recipe pairs, while India has the fewest keywords and the least number of pairs. This disparity is influenced by the number of dish titles provided by Wikipedia and the corresponding number of successfully crawled image-recipe pairs.

Culture	Query Keywords	Crawled Pairs	Percentage (%)
Indonesia	310	24,766	24
Malaysia	112	18,848	18
Thailand	182	23,781	22
Vietnam	102	20,846	20
India	70	16,601	16

The following two tables present the overlap percentages (%) in ingredients and actions, respectively. As shown, Indonesia and Malaysia share a high degree of overlap in both ingredients and cooking actions, while India exhibits the least overlap with other cultures in both categories.

Pairwise ingredient overlap percentages (%) between cultures

	Indonesia	Malaysia	Thailand	Vietnam	India
Indonesia	100	39	31	36	21
Malaysia	39	100	26	29	23
Thailand	31	26	100	31	13
Vietnam	36	29	31	100	18
India	21	23	13	18	100

Pairwise action overlap percentages (%) between cultures

	Indonesia	Malaysia	Thailand	Vietnam	India
Indonesia	100	47	35	34	31
Malaysia	47	100	43	31	31
Thailand	35	43	100	35	32
Vietnam	34	31	35	100	29
India	31	32	32	29	100

However, due to copyright restrictions, we are unable to directly release the food images. Instead, we provide recipe links, allowing researchers to download the corresponding images for further study.