Photo search appeared on Megamarket

2023: Realizing the ability to find things from photos based on the Layer platform

Finding your favorite clothes, shoes and accessories on the Megamarket will now help search by photos. Users can upload both photos of individual items and a whole image - the neural network will select the most similar products and offer them with a list. The company announced this on November 7, 2023.

Photo search works thanks to computer vision technology, which was created on the basis of the Layer platform from Sberbank. The neural network studied for more than a year and a half - during this time millions of photos from fashion magazines, from design shows and photo shoots passed through it.

We can say that photo search is becoming a kind of standard in the Fashion category. With its help, you can find both individual wardrobe items and collect a full bow - for example, following the example of a popular style, the image of your favorite star or a specific model from the show. So, you can take a photo from a fashion magazine and in a few moments pick up the most similar products - put them in the basket and order. The accuracy of the technology is very high - according to testing data, the accuracy reaches 98%, - said Igor Rozhkov, Marketing Director of Megamarket.

The neural network recognizes not only clothes, but also gender, as well as the age of the person in the picture. This was made possible by the implementation of Layer's own, visual transformer-based MiVOLO model, which greatly simplifies search and allows you to make more accurate recommendations. For example, the neural network will not offer dresses for adults if the photo shows a child, and will also offer men's wardrobe items if a man is in the picture.

With advanced algorithms, MiVOLO outperforms many other models. So, she will easily cope with the determination of gender and assume age even if there is no face in the frame or the model stands with its back to the camera. The neural network copes better than a person: as tests show, when viewing photos, respondents on average allow an error of up to 7 years when determining age, while algorithms - up to 4 years.

You can download jpg, jpeg and png images into the search, while there are no restrictions on the size of the photo. The issue takes place by the greatest match, and if necessary, all goods can be easily filtered by category.

Since the neural network trained in a variety of photographs, the background does not matter - it can even be a simple picture where a person is on the street or among various attributes for a photo shoot. It is worth considering that if you need to find a specific accessory, for example, glasses or a bag, then it is better to choose photos where they are completely in the frame and close-up, otherwise the neural network will select all things from the picture as recommendations, "added Igor Rozhkov, Marketing Director of Megamarket.