Wizart finishing materials' solution is based on the neural network sets also other machine learning algorithms. The quality of data used in the model is crucial to ensure high indicators. We are trying to support a variety of finishing materials and decors in terms of visualization at homes. It requires constant analysis and current algorithms and datasets improvement.
It is hard to imagine the diversity of interiors we’ve received from all over the world. And Wizart visualizer needs to process them with the same quality. One of the issues, in this case, is the balance of data according to the room type.
Regarding the finishing material type that is tried on (tiles, wallpaper, paint, etc.), our team worked with entirely different types of space. It can be the kitchen, children's room, office, bedroom, bathroom, and so on.
Each type has its unique sets of furniture, finishing materials, and other parameters, which complicates the development of universal algorithms that work equally well in all those cases.
After analyzing the received data, we noticed that the quantitative distribution by interior types is entirely different. Some groups can significantly dominate others, and this disproportion also negatively affects the work of algorithms when processing photos with less popular interior types.
To solve this problem, we have developed a neural network that can automatically determine the interior type from the picture and do it quite accurately. It allows programming to create certain logic depending on the neural network.
During the experiments, we identified that the main types of interiors are the following:
Now the model automates the type of interior with an accuracy of more than 80%. Thanks to this network, we are capable of automating the analysis of data according to these criteria and balancing the data to further improve the rest of the machine learning solutions. Thus, we plan to get more control over the quality of our visualizer when processing photos, regardless of the type of room.
This is a useful solution for closing our internal needs. And what if not only for our internal needs? We tried to turn everything around and concluded that this neural network can be used for the existing solutions of our clients improvement.
This problem has become urgent for us with the support of an increasing variety of different finishing and decorative materials. But there is a downside, this is the issue of finding suitable finishing materials for your room in the open spaces of an online store. In the presence of large catalogs with a vast number of product categories, the search can become exhausting for the client, its focus will be scattered on product categories that are not suitable for his room.
This can be avoided using our network by implementing business logic and offering your customers more relevant products depending on which room the user selects them for. Perhaps it will improve user experience and conversions, or you will find your own, other use for our network.
In this regard, we decided to make this network publicly available and opened access to it through our Vision API. In the documentation you can see examples and query parameters for this network, or you can use our JavaScript package for faster integration into your store.