Chinfood1000: A large benchmark dataset for chinese food recognition

Abstract

In this paper, we introduce an 1000-category food dataset Chin-Food1000 and propose a simple and effective baseline approach. To our best knowledge, the proposed ChinFood1000 dataset enjoys the largest number of food categories among all publicly available food dataset currently. The categories of the ChinFood1000 dataset are carefully selected to include the most popular Chinese dishes. The dataset includes 852 categories of Chinese dishes, together with 91 classes of drinks and snacks, 26 kinds of fruits and 31 kinds of other food. The images in the dataset present both large inter-class affinity and high intra-class variance. To illustrate the challenges presented by the dataset, a baseline based on a very deep CNN is proposed. In the experiments, the baseline approach is evaluated on three most widely used food datasets and achieves the best performance on all of them. The baseline approach is also applied to the ChinFood1000 dataset, with a promising accuracy reported.