VADET: Visual Attributes Discovery and Exploration Tool

VADET Help

VADET (Visual Attributes Discovery and Exploration Tool) is a demo of a product exploration engine suitable for fashion e-shops and similar enterprises. VADET tool provides some common e-shop functionalities such as content-based category hierarchy, displaying detail of an object, add to cart or proceed to checkout. Furthermore, VADET implements multi-example query search based on visual image descriptors, organizes results into the outfit recommendation and a component for latent visual attributes discovery.

Above, you can see a labeled screenshot of the VADET tool. Content-based category hierarchy and stored example queries are displayed on the left panel. Visual search parameters, queried objects and top-k results are in the central area. Outfit recommendations as well as the content of the shopping cart and administrator tools for visual attributes definition are present on the right sliding panels.

Multi-example queries

Users can run multi-example visual queries by checking the checkboxes (below the product's image) of selected objects and click on "Run search" button.

Multi-example queries are evaluated in two steps. First, the individual distances between the descriptors of query members and candidate objects are computed and individual distances are thereafter aggregated for all query members. The choice of image descriptor is performed by the user through the slider "Level of Details" as well as the selection of aggregation method.

Outfit Recommendation Component

While the query results are displayed to the user directly in the central panel, VADET also implements the Outfit Recommendation component aiming to provide more insight on the query result. Outfit recommendation is based on the relational attributes of the dataset and application of rules from a knowledge-based recommender system.

Visual Attributes Discovery Component

Visual Attributes Discovery (VAD) component is a demo of an administrator's tool linked with the exploration model. Based on the visual queries, administrators can define novel visual attributes, semi-automatically select objects which posses the desired attribute and thus enhance the source relational dataset. Currently, VAD component relies on queries supplied by the administrator and binary relevance . In the production version, this functionality should be extended by the queries based on user exploration sequences or shopping cart content. Administrator's task would be to confirm that relevant visual attributes appear in these sequences and that the fuzzy relevance can be approximated by the selected visual similarity metric.