Generating Knowledge-Enriched Image Annotations for Fine-Grained Visual Classification

Exploiting high-level visual knowledge is the key for a great leap in image classification, in particular, and computer vision, in general. In this paper, we present a tool for generating knowledge-enriched visual annotations and use it to build a benchmarking dataset for a complex classification problem that cannot be solved by learning low and middle-level visual descriptor distributions only. The resulting VegImage dataset contains 3,872 images of 24 fruit varieties, over than 60,000 bounding boxes (portraying the different varieties of fruits as well as context objects such as leaves, etc.) and a large knowledge base (over 1,000,000 OWL triples) containing a-priori knowledge about object visual appearance. We also tested existing fine-grained and CNN-based classification methods on this dataset, showing the difficulty of purely visual-based methods in tackling it.