Abstrakt
When searching for complex data entities, such as products in an e-shop, relational attributes are used as filters within structured queries. However, in many domains the visual appearance of an item is important for a user, while coverage of visual appearance by relational attributes is left to database designer at design time and is by nature an incomplete and imperfect representation of the entity.
Recent advances in computer vision, dominated by deep convolutional neural networks (DCNNs), are a promising tool to cover the gaps. It has been shown that activations of neurons of DCNNs correspond to understandable visual-semantic features of an input image.
We envision that activations of neurons are of great use for search queries in domains with strong visual information, even when obtained from DCNNs models pre-trained on general imagery. Locally scoped visual features obtained using them can be combined to form search masks which would correlate to what humans understand as an attribute, when applied on the entire dataset.
Ultimately, combination of visual features can be identified automatically and formed into immediate suggestion of a new relational attribute, leaving one last task for humans to turn this into augmentation of the database schema - putting a label on it.