DAISO-100

The dataset consists of 160,000 images of 100 class sundries (miscellaneous products). A total of 80,000 images are available for training and testing. The classification task for 100 classes is considerably more difficult than that for 50 classes, because each product has a similar confusing product. Each image involves meta-data regarding the environment: illumination conditions, camera angle, and product conditions. Artificial domain shifts can be created using these meta-data.

该数据集包含160000张图像，涵盖100类杂项商品（sundries），其中80000张图像可用于训练与测试。100类的分类任务难度显著高于50类分类任务，原因在于每一款产品均存在与之相似的易混淆竞品。每张图像均附带环境元数据（meta-data），涵盖光照条件、拍摄角度以及产品状态三类信息。可借助这些元数据构建人工域偏移（domain shift）。