The Empora Fashion Search Engine

On Empora.com you can browse through your favourite stores and brands at the same time narrowing your search down by the styles and colours that catch your eye. No description required - just point and click on the "MORE LIKE THIS" button. Clicking "MORE LIKE THIS" performs a similarity search and our unique technology returns results based on the features of the image you selected. You can navigate this way to view all similar products from different stores and brands, and even refine it by your price range and preferred colour.

This solves the age-old problem of shopping online which meant visiting all of your favourite stores one at a time to see and compare your choices. It also eliminates the frustrations of describing key words for search boxes, which was time-consuming and never really gave you satisfactory results.

The Empora way of browsing saves you time and effort to find what you are after, not to mention - it is much more fun!

How we do it (the technical stuff)

Our objective is to make the visual content of the World Wide Web searchable in more intuitive ways. Fashion is our starting point. The techniques we draw on in order to achieve this come from the fields of machine learning, image processing and computer vision, three vibrant research disciplines with a plethora of interesting, new developments. The idea is simple: instead of relying on tags that may or may not have been added to the images, we try to understand what the image depicts based on a pixel-wise analysis. Colour and texture properties are extracted as well as particular objects. This automatically derived knowledge can subsequently be exploited to help users query and navigate more efficiently through potentially quite large collections of fashion items.

For example, you can select any image and view others that are visually similar under different aspects. The way textual and visual data are combined should make it much easier to find your dream dress, and along with it a perfectly matching bag.

Research papers (from the brains behind the technology)

Our research group is led by Dr Heesch, one of the founders. At the core of our activities lies the problem of image understanding. The problem couldn't be more difficult. It has occupied and consumed generations of bright scientists, and will continue to do so for a long time, Empora's best efforts notwithstanding. Particular areas we are interested in are large-scale multi-class classification, new methods for unsupervised object segmentation, and the development of new visual features for shape analysis. The work is a mixture of long-term foundational research and applied problem solving.

A sample of recent publications and talks by group members is given below.

A survey of browsing models for content based image retrieval (2008) Multimedia Tools and Applications 40, 261-284 (Springer) paper at: http://www.daniel-heesch.com/papers/heesch-08a.pdf

Two step relevance feedback for interactive sense disambiguation in image retrieval (2008) International Conference Visual Information Systems paper at: http://www.daniel-heesch.com/papers/heesch-08b.pdf

Image Search for the World Wide Web (2009) Invited Lecture (IET London, 20th Jan) slides at: http://www.daniel-heesch.com/papers/heesch-09a-slides.pdf