We have worked very hard during the past several months in Samsamia. We were working a lot on improving the accuracy of our visual search system. We have minimized the false positive results. If we are looking for a red bag it is unacceptable to obtain a black bag. The hard work has produced the results we were looking for and we can now say that we are very proud of the precision of our search engine.
Identifying a garment in a picture is not an easy task. There are many problems that have to be addressed.
From the point of view of computer vision, first, it is necessary to remove the background and clean the image. If there is an occlusion in the image, such as when a hand is hiding part of the garment, it has to be detected and fixed. Then we perform a first analysis to obtain a signature that takes into account low-level features like color, shape, relative position or texture.
The second part is to perform a high-level analysis. This analysis tries to imitate the generalization process that appears in the consumer mind when he or she compares several images. The goal is to compose a high-level classifier for every kind of object. This classifier has to be able to recognize the garment type, the use we usually give to the garment and, what is more difficult, the style or subjective perception. The style differs from one person to another. If a person prefers a classic dress style, we will try to obtain results that fits that style. For that purpose, we included a machine learning process in the system to be able to understand the user style and learn from his choices. The more our system is used the more it learns from the user behaviour.