Dataset Card for FoodSeg103

Dataset Summary

FoodSeg103 is a large-scale benchmark for food image segmentation. It contains 103 food categories and 7118 images with ingredient level pixel-wise annotations. The dataset is a curated sample from Recipe1M and annotated and refined by human annotators. The dataset is split into 2 subsets: training set, validation set. The training set contains 4983 images and the validation set contains 2135 images.

Supported Tasks and Leaderboards

No leaderboard is available for this dataset at the moment.

Dataset Structure

Data categories

id	ingridient
0	background
1	candy
2	egg tart
3	french fries
4	chocolate
5	biscuit
6	popcorn
7	pudding
8	ice cream
9	cheese butter
10	cake
11	wine
12	milkshake
13	coffee
14	juice
15	milk
16	tea
17	almond
18	red beans
19	cashew
20	dried cranberries
21	soy
22	walnut
23	peanut
24	egg
25	apple
26	date
27	apricot
28	avocado
29	banana
30	strawberry
31	cherry
32	blueberry
33	raspberry
34	mango
35	olives
36	peach
37	lemon
38	pear
39	fig
40	pineapple
41	grape
42	kiwi
43	melon
44	orange
45	watermelon
46	steak
47	pork
48	chicken duck
49	sausage
50	fried meat
51	lamb
52	sauce
53	crab
54	fish
55	shellfish
56	shrimp
57	soup
58	bread
59	corn
60	hamburg
61	pizza
62	hanamaki baozi
63	wonton dumplings
64	pasta
65	noodles
66	rice
67	pie
68	tofu
69	eggplant
70	potato
71	garlic
72	cauliflower
73	tomato
74	kelp
75	seaweed
76	spring onion
77	rape
78	ginger
79	okra
80	lettuce
81	pumpkin
82	cucumber
83	white radish
84	carrot
85	asparagus
86	bamboo shoots
87	broccoli
88	celery stick
89	cilantro mint
90	snow peas
91	cabbage
92	bean sprouts
93	onion
94	pepper
95	green beans
96	French beans
97	king oyster mushroom
98	shiitake
99	enoki mushroom
100	oyster mushroom
101	white button mushroom
102	salad
103	other ingredients

Data Splits

This dataset only contains two splits. A training split and a validation split with 4983 and 2135 images respectively.

Dataset Creation

Curation Rationale

Select images from a large-scale recipe dataset and annotate them with pixel-wise segmentation masks.

Source Data

The dataset is a curated sample from Recipe1M.

Initial Data Collection and Normalization

After selecting the source of the data two more steps were added before image selection.

1. Recipe1M contains 1.5k ingredient categoris, but only the top 124 categories were selected + a 'other' category (further became 103).
2. Images should contain between 2 and 16 ingredients.
3. Ingredients should be visible and easy to annotate.

Which then resulted in 7118 images.

Annotations

Annotation process

Third party annotators were hired to annotate the images respecting the following guidelines:

1. Tag ingredients with appropriate categories.
2. Draw pixel-wise masks for each ingredient.
3. Ignore tiny regions (even if contains ingredients) with area covering less than 5% of the image.

Refinement process

The refinement process implemented the following steps:

1. Correct mislabelled ingredients.
2. Deleting unpopular categories that are assigned to less than 5 images (resulting in 103 categories in the final dataset).
3. Merging visually similar ingredient categories (e.g. orange and citrus)

Who are the annotators?

A third party company that was not mentioned in the paper.

Additional Information

Dataset Curators

Authors of the paper A Large-Scale Benchmark for Food Image Segmentation.

Licensing Information

Apache 2.0 license.

Citation Information

@inproceedings{wu2021foodseg,
    title={A Large-Scale Benchmark for Food Image Segmentation},
    author={Wu, Xiongwei and Fu, Xin and Liu, Ying and Lim, Ee-Peng and Hoi, Steven CH and Sun, Qianru},
    booktitle={Proceedings of ACM international conference on Multimedia},
    year={2021}
}