Central-Asian-Food-Dataset

In this work, we propose the first Central Asia Food Dataset, containing 16,499 images across 42 classes.

The dataset is unbalaced. The statistics across all 42 classes is shown on Figure below.

Download the dataset

The dataset can be downloaded using the link below. If there are some issues with the link, please, email us on [email protected]

https://issai.nu.edu.kz/wp-content/themes/issai-new/data/models/CAFD/CAFD.zip

Pre-trained models

To illustrate the performance of different classification models on CAFD we have trained different models. We used the largest publicly available fine-grained dataset Food1K [1] that contains 1,000 food classes to evaluate the performance of classifier with the 1,042 food categories.

Model	CAFD (Top-1 Acc.)	CAFD (Top-5 Acc.)	Food1K+CAFD (Top-1 Acc.)	Food1K+CAFD (Top-5 Acc.)
VGG-16	86.03	98.33	80.87	96.19
Squeezenet1_0	79.58	97.29	69.16	90.15
ResNet50	88.03	98.44	83.22	97.25
ResNet101	88.51	98.44	84.20	97.45
ResNet152	88.70	98.59	84.75	97.58
ResNext50_32	87.95	98.44	84.81	97.65
Wide ResNet-50	88.21	98.59	85.27	97.81
DenseNet-121	86.95	98.26	82.45	96.93
EfficientNet-b4	81.28	97.37	87.75	98.01

Pre-trained model weights of the best performing models: ResNet152 on KFD and EfficientNet-b4 on Food1K+KFD can be downloaded using these links:

ResNet152 trained on CAFD:

https://issai.nu.edu.kz/wp-content/themes/issai-new/data/models/CAFD/cafd_resnet152.pt

EfficientNet-b4 trained on Food1K+CAFD:

https://issai.nu.edu.kz/wp-content/themes/issai-new/data/models/CAFD/food1k_kfd_efficientnet.pt

Model training and testing

To train and test using pre-trained models use train.py and test.py files.

References

[1] Min, Weiqing and Wang, Zhiling (2021). Large Scale Visual Food Recognition. arXiv.

In case of using our dataset and/or pre-trained models, please cite our work:

@Article{nu15071728,
AUTHOR = {Karabay, Aknur and Bolatov, Arman and Varol, Huseyin Atakan and Chan, Mei-Yen},
TITLE = {A Central Asian Food Dataset for Personalized Dietary Interventions},
JOURNAL = {Nutrients},
VOLUME = {15},
YEAR = {2023},
NUMBER = {7},
ARTICLE-NUMBER = {1728},
URL = {https://www.mdpi.com/2072-6643/15/7/1728},
ISSN = {2072-6643}
}