FoodSeg103 is a new food image dataset containing 7,118 images (4,983 training images and 2,135 test images) and annotations of 104 ingredient categories, with an average of 6 annotations and pixel-level masks per image. This dataset provides a large-scale benchmark for food image segmentation.

The main challenges of this dataset are:

• High intra-variance of the same food ingredient with different cooking methods
• Long-tail distribution
• Complex contexts