Sai-Chileandiet: A Multi-Label Food Dataset With Self-Acquired Images Of The Chilean Diet

The lack of data representative of the Chilean diet precludes providing accurate applications based on real foods consumed by Chilean society. Moreover, for the multi-label food recognition problem, there are few datasets available, which is mandatory for food analysis considering the great variability of dishes that can be made by combining different foods, which do not have a particular name and therefore cannot be treated as a single-label recognition problem. In this paper, we introduce a new challenging multi-label food dataset, SAI-ChileanDiet, focused on the Chilean diet. The dataset includes foods of popular dishes from Chilean cuisine and others included in the Chilean diets. SAI-ChileanDiet considers different sources of data for training and test sets. The training set comprises images from the web and the test set comprises self-acquired images. This feature of the dataset allows us to assess the representativeness of public images in real-world Chilean foods and encourages the generation of advanced methods capable of dealing with the distribution shift of these sets. We evaluate on the proposed SAI-ChileanDiet dataset the performance of several state-of-the-art deep networks and AI platforms for multi-label food recognition. The experiments demonstrate how challenging the dataset is and the importance of validating self-acquired data to be aware of the performance it could offer in real-life conditions.