Toward Chinese Food Understanding: a Cross-Modal Ingredient-Level Benchmark

Lanjun Wang¹ Chenyu Zhang¹ An-An Liu¹ Bo Yang¹ Mingwang Hu¹ Xinran Qiao¹ Lei Wang² Jianlin He² Qiang Liu²

¹Tianjin University ²Meituan Group

Content

I. Introduction

This is the supplementary material for the paper "Toward Chinese Food Understanding: a Cross-Modal Ingredient-Level Benchmark" [link]. The web intends to release the proposed dataset CMIngre and introduce related tasks.

II. CMIngre Dataset

CMIngre consists of 8,001 food images with 429 ingredient labels from three sources, where 1,719 from dishes, 2,330 from recipes, and 3,952 from user-generated content (UGC). In Sec. II-A, we introduce the three data sources of CMIngre. In Sec. II-B, we introduce the ingredient ontology [link] for refining ingredient labels. In Sec. II-C, we introduce two visualization tools for classifying and fusing ingredient labels. In Sec. II-D, we compare CMIngre with existing food datasets. In Sec. II-E, we provide a download method of CMIngre.

A. Data Sources

To gather a comprehensive collection of food images, we explore three types of image-text pairings:

Dish Images. As depicted in Figure 1, second row, this category includes images of dishes paired with their names. The text in this type provides the most succinct description compared to the others.
Recipe Images. Shown in Figure 1, third row, these data consist of recipe images accompanied by detailed recipe text. These images are of higher quality and are more informatively described than those in the other two categories.
User-Generated Content (UGC). This type, illustrated in the last row of Figure 1, involves images taken by users and their accompanying comments. As the user-generated content lacks constraint, both images and text descriptions often include elements irrelevant to food, such as restaurant ambiance or tableware.

In Figure 1, we extract ingredient labels from both text annotation and image annotation.

Figure 1. Food images in CMIngre comes from three sources, where UGC refers to the user-generated content such as the user comment.

B. Ingredient Ontology

Since some ingredient labels of different names referring to the same ingredient, for example, "松花蛋–preserved egg" and "皮蛋–preserved egg", we utilize an ingredient ontology from the People’s Republic of China health industry standard [link] to compare and combine the ingredient labels. In Figure 2, we show the complete sub-tree under the super-class (i.e. the second level) "Dried beans and products", where the leaf nodes are ingredient labels after cleaning up and the non-leaf nodes are from the standard.

Figure 2. The ontology used to clean labels in CMIngre. In this figure, we show the complete sub-tree under the super-class (i.e. the second level) "Dried beans and products". The leaf nodes are ingredient labels after cleaning up and the non-leaf nodes are from the standard.

C. Visualization Tools

In order to categorize ingredient labels into the ingredient ontology, we have designed a classification tool (provided in the "Label_Classification" folder). Then, we have developed a fusing tool (provided in the "Label_Fusion" folder) to merge ingredients with identical semantics under the same parent node in the ingredient ontology.

D. Comparison with Existing Food Datasets

We compare CMIngre with other food related datasets in Table 1. It can be observed that existing food-related datasets mainly focus on the food recognition task, which aims to recognize the food category within the image. Although few datasets do provide annotations for food bounding boxes, their objective is to locate the entire dish, not the free-form ingredients. In contrast, Recipe 1M offers ingredient annotations for each food image. However, due to the absence of location annotations for these fine-grained ingredients, they only implicitly model the associations between entire food images and corresponding ingredients, limiting the model performance. Consequently, we introduce CMIngre, aimed at enhancing the understanding of Chinese food by ingredient detection and retrieval tasks.

Table 1. The statistical comparison between existing food-related datasets and CMIngre.

Dataset	Task	Image Number	Annotation Category	Number of Annotation Category	BBox
ChileanFood64	Food Recognition	11,504	Food	64	✓
UECFood256	Food Recognition	29,774	Food	256	✓
UNIMIB2016	Food Recognition	1,027	Food	73	✓
ISIA Food-500	Food Recognition	1,027	Food	73	✕
Food2K	Food Recognition	1,036,564	Food	2,000	✕
Recipe 1M	Recipe Retrieval	1,029,720	Recipe	1,047	✕
CMIngre	Ingredient Detection & Retrieval	8,001	Ingredient	429	✓

E. Download

You can download CMIngre dataset from HuggingFace [Dataset Card].

III. Task

Our dataset involves two tasks, i.e. ingredient detection and cross-modal ingredient retrieval.

Ingredient detection focuses on identifying the ingredients and providing precise location information within the image. As shown in Figure 3, we locate and identify the ingredients in food images from dish, recipe, and UGC.

Figure 3. The ingredient detection visualization of theree food images from dish, recipe, and UGC.

Cross-modal ingredient retrieval aims to investigate the intricate relationship between the image and the composition of ingredients. We visualize top-5 retrieval results by randomly sampling a query object from dish, recipe, and UGC in the test set. As shown in Figure 4, the corresponding ingredient composition appears in the first index position of the retrieval list with the highest matching similarity. Similarly, as shown in Figure 5, the corresponding image appears in the first index position of the retrieval list with the highest matching similarity.

Figure 4. The top-5 retrieval visualization of three random query images from dish, recipe, and UGC.

Figure 5. The top-5 retrieval visualization of three query ingredient composition from different sources.

IV. Citation

BibTeX:
@inproceedings{li2023photomaker, title={Toward Chinese Food Understanding: a Cross-Modal Ingredient-Level Benchmark}, author={Wang, Lanjun and Zhang, Chenyu and Liu, An-An and Yang, Bo and Hu, Mingwang and Qiao, Xinran and Wang, Lei and He, Jianlin and Liu, Qiang}, booktitle={IEEE Transactions on Multimedia}, year={2024}, publisher={IEEE} }

image imagewidth (px) 29 6.87k